Alternatives to Patronus AI

GLM-5.3 is Z.ai’s advanced coding and agentic reasoning model built through scaled post-training on top of the GLM-5.2 base model. The release focuses on frontier coding, long-horizon software engineering, agent tasks, cyber evaluation, and reinforcement learning at scale. GLM-5.3 improves significantly over GLM-5.2 on complex coding benchmarks, real-world engineering environments, Terminal Bench 3.0, DeepSWE, Agents’ Last Exam, and Z.ai’s internal Code Bench. The model is trained on environments that resemble real professional work, including tasks involving codebases, infrastructure, documentation, compute clusters, experiments, bottleneck diagnosis, implementation, testing, and measurable optimization. Z.ai’s post-training stack includes IndexShare for efficient long-context processing, SAO for reinforcement learning on long-horizon tasks, and slime for large-scale asynchronous training. GLM-5.3 supports three thinking effort levels, including low, high, and max, with max recommended for coding tasks. The model also demonstrates emergent cyber capabilities across vulnerability discovery and exploitation benchmarks, prompting continued safety evaluation and hardening before weights are released. GLM-5.3 can be used through the GLM Coding Plan, ZCode, Claude Code, OpenCode, and other coding agent workflows. By combining stronger coding performance, long-horizon task execution, post-training scale, cyber evaluation, reasoning effort controls, and coding-agent integrations, GLM-5.3 supports advanced developer and research workflows.

MiMo-V2.6-Pro is an open-source, natively omnimodal AI model from Xiaomi MiMo designed for coding, agentic workflows, visual creation, research, and computer-based tasks. It is the highest-capability model in the MiMo-V2.6 series and was trained using a large-scale reinforcement learning process focused on verifiable, complex tasks. The model can handle software engineering, automation, tool use, visual reasoning, and long-running agent workflows across multiple environments. Its multimodal capabilities extend beyond traditional coding into 3D scene generation, Blender modeling, embodied simulation, frontend development, presentation design, video creation, and music composition. MiMo-V2.6-Pro can take text, images, video, and other visual inputs and coordinate multi-agent workflows to produce and iteratively refine complex outputs. Research applications demonstrated by Xiaomi include materials discovery, literature analysis, computational simulation, hypothesis generation, and formalizing mathematical proofs in Lean. Xiaomi has released the model alongside its technical report, reinforcement learning environments, and RL code so researchers can study and reproduce the training approach. MiMo-V2.6-Pro is also available in an UltraSpeed configuration that provides substantially faster output for latency-sensitive workflows. Users can access the model through MiMo Desktop, AI Studio, MiMo Code, the MiMo API Platform, OpenRouter, and Hugging Face.

Muse Spark 1.2 is Meta’s newest coding-focused model, released alongside Muse Code as part of Meta’s AI developer platform. The model improves on Muse Spark 1.1 with stronger code generation, complex debugging, codebase understanding, and full developer workflow performance. Muse Spark 1.2 powers Muse Code, a terminal coding agent that can plan changes, write code, validate results, and coordinate persistent background subagents. The model was co-trained with Muse Code so it performs well inside the agentic coding runtime and tool environment. Its training included scaled coding compute, broader training environment diversity, rejection-sampled harness trajectories, recipe optimizations, and Muse Code toolset integration. Muse Spark 1.2 is designed for long-horizon coding tasks such as whole-repository generation, large end-to-end projects, auto-research, and extended optimization work. It uses planning to sequence work, goal conditioning to stay aligned with the user’s objective, and context compaction to preserve useful knowledge over long sessions. The model also benefits from a self-improvement loop where Muse Spark 1.1 generated challenging coding environments and instruction-following templates for training. By combining coding specialization, agentic workflow support, long-horizon training, subagent compatibility, and Meta Model API availability, Muse Spark 1.2 helps developers build, debug, and optimize software more effectively.

Cursor has introduced Composer 2.5, a next-generation AI coding assistant built to deliver stronger reasoning, better collaboration, and improved reliability during software development tasks. The upgraded model performs better on long-running coding workflows and can manage complicated instructions with greater consistency than earlier Composer versions. Cursor expanded the training process by scaling compute resources, generating more advanced reinforcement learning environments, and refining behavioral traits that improve the developer experience. One of the key innovations in Composer 2.5 is its targeted textual feedback system, which helps the model learn from localized mistakes inside long coding trajectories instead of relying only on broad reward signals. This training method allows the AI to improve coding style, communication quality, and tool usage accuracy in a more focused way. The company also increased the amount of synthetic coding data by 25 times compared to Composer 2, giving the model exposure to more difficult and realistic programming tasks. During development, the system demonstrated sophisticated reasoning abilities by uncovering hidden implementation details and reverse-engineering deleted functionality inside synthetic environments. Composer 2.5 additionally uses advanced distributed training methods such as Sharded Muon and dual mesh HSDP to optimize large-scale model training performance. Available directly inside Cursor, the model comes in both standard and fast variants with different pricing tiers designed for developers, teams, and enterprise-scale engineering workflows.

GLM-5 is a next-generation open-source foundation model from Z.ai designed to push the boundaries of agentic engineering and complex task execution. Compared to earlier versions, it significantly expands parameter count and training data, while introducing DeepSeek Sparse Attention to optimize inference efficiency. The model leverages a novel asynchronous reinforcement learning framework called slime, which enhances training throughput and enables more effective post-training alignment. GLM-5 delivers leading performance among open-source models in reasoning, coding, and general agent benchmarks, with strong results on SWE-bench, BrowseComp, and Vending Bench 2. Its ability to manage long-horizon simulations highlights advanced planning, resource allocation, and operational decision-making skills. Beyond benchmark performance, GLM-5 supports real-world productivity by generating fully formatted documents such as .docx, .pdf, and .xlsx files. It integrates with coding agents like Claude Code and OpenClaw, enabling cross-application automation and collaborative agent workflows. Developers can access GLM-5 via Z.ai’s API, deploy it locally with frameworks like vLLM or SGLang, or use it through an interactive GUI environment. The model is released under the MIT License, encouraging broad experimentation and adoption. Overall, GLM-5 represents a major step toward practical, work-oriented AI systems that move beyond chat into full task execution.

Fugu Cyber is an advanced orchestration model designed specifically for contemporary cyber defense, operating as a unified entity through a single API endpoint while adeptly managing multiple specialized agents to tackle intricate security challenges. This innovative model does not rely on a single provider and is tailored for two main defense operations: assessing complex codebases to identify genuine vulnerabilities and converting raw cyber threat intelligence into actionable detection rules. Its performance on CyberGym, which tests vulnerability analysis and validation, resulted in an impressive success rate of 86.9%, whereas on CTI-REALM, which evaluates the generation of detection rules from threat intelligence reports, it achieved a score of 72.1%. These results position Fugu Cyber among the top-tier models focused on cybersecurity innovations. Rather than functioning as an isolated tool, Fugu Cyber is designed to serve as the cognitive engine within larger security infrastructures, enhancing overall defense capabilities against evolving cyber threats. This integration allows for a more holistic approach to cyber defense, enabling organizations to respond more effectively to potential attacks.

LongCat-2.0 represents a significant advancement in the realm of language models, featuring a staggering 1.6 trillion parameters through a Mixture-of-Experts architecture that leverages AI ASIC superpods, with approximately 48 billion parameters engaged per token, showcasing exceptional capabilities in coding and agentic tasks. This model marks a notable improvement over its predecessors by integrating a large-scale sparse architecture with specialized post-training methods tailored for tasks in real-world software development, tool utilization, long-context reasoning, and complex agent workflows. Entirely developed and executed on AI ASIC superpods, LongCat-2.0 underwent pretraining that encompassed over 35 trillion tokens and millions of accelerator hours, exemplifying cutting-edge training methodologies on innovative hardware solutions. To enhance its performance on tasks requiring long-term context, the model incorporates LongCat Sparse Attention and is trained using hundreds of billions of tokens from 1M-context datasets, enabling it to effectively manage ultra-long context tasks and ensure robust understanding of lengthy documents. This combination of features positions LongCat-2.0 as a pioneering force in the landscape of advanced language models.

Prime Intellect serves as a comprehensive superintelligence framework, offering a cohesive platform for computation, training, inference, and experimentation for groups aiming to develop, implement, and enhance their models over time. Instead of relying on advancements from frontier models, the stack emphasizes ownership of intelligence, providing users with a singular loop for reinforcement learning environments, extensive training, evaluations, inference, and computing needs. Within the Lab, teams can enable self-improving agents by transforming tasks into reinforcement learning settings and utilizing the Prime CLI for creation, development, evaluation, and deployment. The Environment Hub presents access to an extensive collection of over 2,500 open-source RL environments, while hosted evaluations allow teams to assess model performance across various open-source frameworks without the burden of managing infrastructure. Additionally, Hosted Training facilitates large-scale models tailored for agentic workflows, ensuring managed training processes with complete visibility and control, along with direct assistance from the dedicated applied research team, allowing for a more robust and user-friendly experience in model development. This integrated approach not only streamlines the development process but also fosters innovation and collaboration among teams.

SWE-1.7 is Cognition’s most capable software engineering model, built to push frontier coding performance while reducing the cost of high-quality agentic rollouts. The model is designed for real-world software development tasks that require extended reasoning, codebase understanding, terminal use, debugging, feature work, migrations, and careful validation. It was trained from a Kimi K2.7 base and improved through Cognition’s reinforcement learning pipeline, including more stable training, stronger infrastructure, better data curation, and long-horizon task techniques. SWE-1.7 is especially optimized for asynchronous software engineering, where an agent needs to work through large projects over longer sessions instead of simply answering short prompts. Its self-compaction capabilities allow the model to summarize its working state and resume from that summary, helping it operate beyond the raw context window on multi-hour tasks. The model is also trained to balance task success with efficiency, using concise reasoning when possible while preserving deeper exploration for harder problems. SWE-1.7 tends to investigate codebases more thoroughly than its base model, reading files, running searches, probing edge cases, and experimenting before making changes. It is available in Devin through web, desktop, and CLI interfaces, with Cerebras serving support at 1000 TPS. SWE-1.7 gives developers and engineering teams a high-performance coding model for complex software projects at a more practical cost.

MiniMax M2.5 is a next-generation foundation model built to power complex, economically valuable tasks with speed and cost efficiency. Trained using large-scale reinforcement learning across hundreds of thousands of real-world task environments, it excels in coding, tool use, search, and professional office workflows. In programming benchmarks such as SWE-Bench Verified and Multi-SWE-Bench, M2.5 reaches state-of-the-art levels while demonstrating improved multilingual coding performance. The model exhibits architect-level reasoning, planning system structure and feature decomposition before writing code. With throughput speeds of up to 100 tokens per second, it completes complex evaluations significantly faster than earlier versions. Reinforcement learning optimizations enable more precise search rounds and fewer reasoning steps, improving overall efficiency. M2.5 is available in two variants—standard and Lightning—offering identical capabilities with different speed configurations. Pricing is designed to be dramatically lower than competing frontier models, reducing cost barriers for large-scale agent deployment. Integrated into MiniMax Agent, the model supports advanced office skills including Word formatting, Excel financial modeling, and PowerPoint editing. By combining high performance, efficiency, and affordability, MiniMax M2.5 aims to make agent-powered productivity accessible at scale.

AfterQuery serves as a practical research platform aimed at generating high-quality training datasets for cutting-edge artificial intelligence models by emulating the cognitive processes of seasoned professionals as they think, reason, and tackle challenges in their fields. By converting real-world work scenarios into organized datasets, it provides insights that transcend mere outputs, incorporating intricate decision-making, trade-offs, and contextual reasoning that typical internet-sourced data fails to capture. The platform collaborates closely with subject matter experts to produce supervised fine-tuning data, which includes prompt–response pairs alongside comprehensive reasoning trails, in addition to reinforcement learning datasets featuring expertly crafted prompts and assessment frameworks that translate subjective evaluations into scalable reward mechanisms. Furthermore, it develops customized agent environments using various APIs and tools, facilitating the training and evaluation of models within realistic workflows while also tracking computer-use trajectories that illustrate how individuals engage with software in a detailed, step-by-step manner. This multi-faceted approach ensures that the data generated not only reflects expert insights but is also adaptable for a wide range of applications in the evolving landscape of artificial intelligence.

micro1 Intelligence is a data research company focused on advancing frontier artificial intelligence through expert human data, contextual evaluations, and realistic training environments. The company develops infrastructure that enables AI models and autonomous agents to learn from high-quality human expertise rather than relying solely on synthetic or conventional datasets. Its Realm platform creates reinforcement learning environments that simulate real-world scenarios to generate human feedback and improve agent reasoning capabilities. Cortex provides contextual evaluation tools that measure and optimize AI agent performance in production environments using realistic tasks and benchmarks. micro1 also develops robotics datasets that capture high-fidelity real-world interactions to support the training of next-generation embodied AI systems. Alongside its technology platforms, the company publishes research on AI benchmarking, extraction systems, pathology reasoning, human data markets, and model evaluation. Expert opportunities and data partnerships allow professionals and organizations to contribute specialized knowledge that improves AI training quality. By combining research, human expertise, and production-focused evaluation, micro1 Intelligence helps accelerate the development of safer, more capable AI systems. The platform is designed to support frontier AI companies building intelligent agents, reasoning systems, and robotics applications with higher-quality training data.

Qwen3-Coder is an advanced code model that comes in various sizes, prominently featuring the 480B-parameter Mixture-of-Experts version (with 35B active) that inherently accommodates 256K-token contexts, which can be extended to 1M, and demonstrates cutting-edge performance in Agentic Coding, Browser-Use, and Tool-Use activities, rivaling Claude Sonnet 4. With a pre-training phase utilizing 7.5 trillion tokens (70% of which are code) and synthetic data refined through Qwen2.5-Coder, it enhances both coding skills and general capabilities, while its post-training phase leverages extensive execution-driven reinforcement learning across 20,000 parallel environments to excel in multi-turn software engineering challenges like SWE-Bench Verified without the need for test-time scaling. Additionally, the open-source Qwen Code CLI, derived from Gemini Code, allows for the deployment of Qwen3-Coder in agentic workflows through tailored prompts and function calling protocols, facilitating smooth integration with platforms such as Node.js and OpenAI SDKs. This combination of robust features and flexible accessibility positions Qwen3-Coder as an essential tool for developers seeking to optimize their coding tasks and workflows.

Qwen3-Coder is a versatile coding model that comes in various sizes, prominently featuring the 480B-parameter Mixture-of-Experts version with 35B active parameters, which naturally accommodates 256K-token contexts that can be extended to 1M tokens. This model achieves impressive performance that rivals Claude Sonnet 4, having undergone pre-training on 7.5 trillion tokens, with 70% of that being code, and utilizing synthetic data refined through Qwen2.5-Coder to enhance both coding skills and overall capabilities. Furthermore, the model benefits from post-training techniques that leverage extensive, execution-guided reinforcement learning, which facilitates the generation of diverse test cases across 20,000 parallel environments, thereby excelling in multi-turn software engineering tasks such as SWE-Bench Verified without needing test-time scaling. In addition to the model itself, the open-source Qwen Code CLI, derived from Gemini Code, empowers users to deploy Qwen3-Coder in dynamic workflows with tailored prompts and function calling protocols, while also offering smooth integration with Node.js, OpenAI SDKs, and environment variables. This comprehensive ecosystem supports developers in optimizing their coding projects effectively and efficiently.

Cognition has unveiled SWE-1.5, the newest agent-model specifically designed for software engineering, featuring an expansive "frontier-size" architecture composed of hundreds of billions of parameters and an end-to-end optimization (encompassing the model, inference engine, and agent harness) that enhances both speed and intelligence. This model showcases nearly state-of-the-art coding capabilities and establishes a new standard for latency, achieving inference speeds of up to 950 tokens per second, which is approximately six times quicker than its predecessor, Haiku 4.5, and thirteen times faster than Sonnet 4.5. Trained through extensive reinforcement learning in realistic coding-agent environments that incorporate multi-turn workflows, unit tests, and quality assessments, SWE-1.5 also leverages integrated software tools and high-performance hardware, including thousands of GB200 NVL72 chips paired with a custom hypervisor infrastructure. Furthermore, its innovative architecture allows for more effective handling of complex coding tasks and improves overall productivity for software development teams. This combination of speed, efficiency, and intelligent design positions SWE-1.5 as a game changer in the realm of coding models.

Grok 4.1 Fast represents xAI’s leap forward in building highly capable agents that rely heavily on tool calling, long-context reasoning, and real-time information retrieval. It supports a robust 2-million-token window, enabling long-form planning, deep research, and multi-step workflows without degradation. Through extensive RL training and exposure to diverse tool ecosystems, the model performs exceptionally well on demanding benchmarks like τ²-bench Telecom. When paired with the Agent Tools API, it can autonomously browse the web, search X posts, execute Python code, and retrieve documents, eliminating the need for developers to manage external infrastructure. It is engineered to maintain intelligence across multi-turn conversations, making it ideal for enterprise tasks that require continuous context. Its benchmark accuracy on tool-calling and function-calling tasks clearly surpasses competing models in speed, cost, and reliability. Developers can leverage these strengths to build agents that automate customer support, perform real-time analysis, and execute complex domain-specific tasks. With its performance, low pricing, and availability on platforms like OpenRouter, Grok 4.1 Fast stands out as a production-ready solution for next-generation AI systems.

Hyta is an innovative platform that facilitates the scaling and operationalization of AI workflows after training by establishing continuous, always-on pipelines that combine specialized human intelligence with a focus on monitoring reliable contributions, ensuring that model enhancement is an ongoing endeavor instead of a singular effort. This platform brings together a collective of domain experts and machine-learning collaborators who provide valuable human insights essential for long-term, domain-specific model training and reinforcement learning frameworks, while also implementing strategies to maintain contributor trust and context throughout various projects and models. By customizing pipelines to meet the unique requirements of organizations and specific projects, Hyta guarantees dependable progress, safeguards verified contributions, and allows for ongoing feedback, thereby enhancing capabilities across diverse industries. In addition to connecting contributors, research labs, companies, and post-training teams, Hyta fosters a comprehensive ecosystem that empowers organizations to manage human-in-the-loop workflows on a large scale, seamlessly integrating human feedback into the continuous model development process. Furthermore, this interconnected approach not only improves the efficiency of AI models but also enriches the collaboration between human expertise and machine learning, driving innovation and better outcomes in AI applications.

MiniMax M2.7 is a powerful AI model built to drive real-world productivity across coding, search, and office-based workflows. It is trained using reinforcement learning across a wide range of real-world environments, enabling it to execute complex, multi-step tasks with precision and efficiency. The model demonstrates strong problem-solving capabilities by breaking down challenges into structured steps before generating solutions across multiple programming languages. It delivers high-speed performance with rapid token output, ensuring faster completion of demanding tasks. With optimized reasoning, it reduces token usage and execution time, making it more efficient than previous models. M2.7 also achieves state-of-the-art results in software engineering benchmarks, significantly improving response times for technical issues. Its advanced agentic capabilities allow it to work seamlessly with tools and support complex workflows with high skill accuracy. The model is designed to handle professional tasks, including multi-turn interactions and high-quality document editing. It also provides strong support for office productivity, enabling efficient handling of structured data and business tasks. With competitive pricing, it delivers high performance while remaining cost-effective. Overall, it combines speed, intelligence, and versatility to meet the needs of modern professionals and teams.

ERNIE 5.1 is Baidu’s next-generation large language model engineered to provide advanced reasoning, autonomous agent capabilities, creative writing performance, and enterprise-grade AI intelligence with highly optimized efficiency. Built on the pre-training foundation of ERNIE 5.0, the model significantly reduces parameter size and computational requirements while still delivering leading performance across major international AI benchmarks. ERNIE 5.1 demonstrates strong capabilities in reasoning, mathematical problem solving, knowledge retrieval, search tasks, and agentic workflows that allow it to handle complex multi-step operations and decision-making scenarios. The platform introduces a fully asynchronous reinforcement learning architecture designed to improve scalability, training efficiency, resource utilization, and long-horizon task stability for large-scale AI development. Baidu also implemented a multi-stage reinforcement learning pipeline that separates expert capability training from unified capability fusion, allowing the model to specialize in areas such as coding, reasoning, search, and conversational intelligence without creating performance conflicts between domains. ERNIE 5.1 supports advanced creative generation with improved emotional understanding, narrative structure control, stylistic adaptability, and contextual awareness for writing-intensive applications. The model performs competitively against leading closed-source global AI systems in knowledge benchmarks, reasoning evaluations, and creative content generation tasks. ERNIE 5.1 is also integrated into creative production platforms, AI storytelling systems, roleplay applications, and agentic AI environments that support content creators and enterprise workflows.

Lumen Outpost represents Cosine’s refined post-trained coding model, evaluated against its foundational model Kimi K2.6, along with GPT-5.5, GPT-5.4, and Gemini 3.1 Pro, specifically focusing on intricate, long-term coding assignments across 13 different programming languages. This model is designed not only for precision in coding but also to enhance key behavioral indicators vital in engineering processes, such as agent initiative, strategic planning, scope management, action coherence, succinct updates, and effective communication. According to Cosine’s benchmark analysis, the specialized post-training significantly elevated the base model's performance, with Lumen Outpost surpassing Kimi K2.6 in tests like Niche-Bench, Slop-Bench, Vibe-Bench, as well as in terms of cost efficiency for successful task completion. In the Niche-Bench assessment, which evaluates niche, legacy, and environmentally constrained programming languages, Lumen Outpost attained a score of 53.9% and excelled or equaled performance in 9 out of the 13 languages evaluated, demonstrating marked improvements particularly in Fortran, ABAP, Java, and Rust. The impressive results symbolize a significant leap in the practical application of coding models in real-world scenarios, underscoring the effectiveness of targeted training methodologies.

NVIDIA Isaac Lab is an open-source robot learning framework that utilizes GPU acceleration and is built upon Isaac Sim, aimed at streamlining and integrating various robotics research processes such as reinforcement learning, imitation learning, and motion planning. By harnessing highly realistic sensor and physics simulations, it enables the effective training of embodied agents and offers a wide range of pre-configured environments that include manipulators, quadrupeds, and humanoids, while supporting over 30 benchmark tasks and seamless integration with well-known RL libraries, including RL Games, Stable Baselines, RSL RL, and SKRL. The design of Isaac Lab is modular and configuration-driven, which allows developers to effortlessly create, adjust, and expand their learning environments; it also provides the ability to gather demonstrations through peripherals like gamepads and keyboards, as well as facilitating the use of custom actuator models to improve sim-to-real transfer processes. Furthermore, the framework is designed to operate effectively in both local and cloud environments, ensuring that compute resources can be scaled flexibly to meet varying demands. This comprehensive approach not only enhances productivity in robotics research but also opens new avenues for innovation in robotic applications.

ReinforceNow serves as a comprehensive platform dedicated to ongoing learning through AI agents, designed to assist teams in deploying, training, and iterating efficiently. Developers are empowered to create AI agents that can be continuously trained using production traffic, or they can opt for Claude Code to configure the setup automatically. The platform manages vital components such as reinforcement learning infrastructure, experiment orchestration, agent versioning, GPU training logic, and telemetry, allowing teams to concentrate on refining agent logic, data collection, and reward systems. With support for rapid LLM fine-tuning using LoRA, high-throughput training capabilities, and extensive compatibility with open-source models including Qwen, DeepSeek, and GPT-OSS, ReinforceNow enhances developers' efficiency. It offers sophisticated telemetry features that help evaluate, monitor, and iterate on AI agent LLM applications, including detailed traces, reward systems, experiment metrics, and training visibility. Teams can tackle extended tasks that require context sizes ranging from 32k to 1 million, create specialized agents for multi-turn interactions and long-duration tasks, and access an array of tools to streamline their reinforcement learning workflows, ultimately fostering innovation in AI development.

Distil Labs enhances AI performance by substituting costly calls to advanced models with tailored small language models designed for specific tasks while ensuring the quality standards are upheld. By monitoring real production traffic and gathering traces from current LLM requests, it constructs an evaluation set to gain insights into actual workload behavior. Following this, the company creates and verifies synthetic training data, aligns the data distribution with the intended workload, and engages in supervised fine-tuning alongside reinforcement learning. The model is then quantized, and an optimized endpoint is established. The outcomes are systematically assessed against the existing model concerning accuracy, latency, and efficiency, providing teams with data to determine when to increase traffic. Ultimately, the OpenAI-compatible endpoint features a specialized small language model, prompt optimization, effective caching, and refined serving tailored for the specific application, ensuring maximum performance. This comprehensive approach allows organizations to maximize the potential of their AI implementations.

ERNIE X1.1 is Baidu’s latest reasoning AI model, designed to raise the bar for accuracy, reliability, and action-oriented intelligence. Compared to ERNIE X1, it delivers a 34.8% boost in factual accuracy, a 12.5% improvement in instruction compliance, and a 9.6% gain in agentic behavior. Benchmarks show that it outperforms DeepSeek R1-0528 and matches the capabilities of advanced models such as GPT-5 and Gemini 2.5 Pro. The model builds upon ERNIE 4.5 with additional mid-training and post-training phases, reinforced by end-to-end reinforcement learning. This approach helps minimize hallucinations while ensuring closer alignment to user intent. The agentic upgrades allow it to plan, make decisions, and execute tasks more effectively than before. Users can access ERNIE X1.1 through ERNIE Bot, Wenxiaoyan, or via API on Baidu’s Qianfan platform. Altogether, the model delivers stronger reasoning capabilities for developers and enterprises that demand high-performance AI.

Odyssey-2 Max is an advanced, real-time world simulation model that transcends conventional generative AI by learning the dynamics of the physical world and facilitating ongoing, interactive settings. As the third iteration in the Odyssey-2 series, it boasts a remarkable increase in scale, featuring three times more parameters and ten times the computational power compared to its predecessor, Odyssey-2 Pro, which fosters new emergent behaviors and enhances the stability and realism of simulations. Crafted to accurately replicate physics, human movement, interactions, and environmental changes in real time, it offers continuous visual output that adapts instantaneously to user commands rather than relying on fixed video clips. In contrast to traditional video models that produce short, predetermined sequences, Odyssey-2 Max enables the creation of extensive simulations that evolve in real time, allowing users to engage with a dynamically unfolding environment. This innovative approach redefines user interaction, making every session unique and immersive as the simulation adapts to each new input.

Lucky Robots is an innovative platform dedicated to robotics simulation that empowers teams to train, assess, and enhance AI models for robots within meticulously crafted virtual environments that closely reflect the nuances of real-world physics, sensors, and interactions. This system facilitates the extensive creation of synthetic training data and allows for swift iterations without the need for physical robots or expensive lab environments. By leveraging cutting-edge simulation technology, it constructs hyper-realistic scenarios, such as kitchens and various terrains, enabling the exploration of diverse edge cases and the generation of millions of labeled episodes to support scalable model learning. This approach not only speeds up development but also significantly cuts costs and minimizes safety risks. Additionally, the platform accommodates natural language control in its simulated environments, provides the flexibility for users to upload their own robot models or select from existing commercial options, and incorporates collaborative tools through LuckyHub for sharing environments and training workflows. As a result, developers can optimize their models more effectively for real-world applications, ultimately enhancing the performance and reliability of their robotic solutions.

NVIDIA Alpamayo 2 Super stands as a pioneering open model tailored for robotaxis and autonomous vehicles, designed to navigate rare and intricate driving scenarios while generating decisions that developers can analyze, verify, and rely upon. Utilizing the foundations of NVIDIA Cosmos 3 Super Reasoner and enhanced through reinforcement learning, it merges commercial accessibility with the ability to handle multiple tasks related to autonomous driving. The model comprehensively analyzes full-surround camera input, integrating perspectives from the front, sides, and rear to adeptly manage lane changes, merges, unprotected turns, and complex intersections. In addressing each driving scenario, it can produce a planned trajectory for the vehicle, a chain-of-causation that elucidates the decision-making process, a meta-action such as yielding or stopping, and reasoning auto-labels for both training and validation purposes, along with visual question-answering outputs anchored in specific image regions. These interconnected outputs facilitate the correlation between the model's observations and the actions it undertakes, thereby enhancing transparency in autonomous decision-making. Additionally, this functionality supports developers in refining and optimizing the model's performance in real-world applications.

Salesforce Koa represents the company's inaugural CRM reasoning model tailored for Agentforce, developed using NVIDIA Nemotron and enhanced by 27 years of accumulated Salesforce CRM expertise to navigate intricate, multi-step tasks in enterprise environments. Rooted in nearly three decades of CRM application experience, it has undergone additional training on a unique synthetic dataset that reflects authentic business processes, workflows, and operational regulations. The scenarios employed during training are designed to replicate the reasoning, tool utilization, and decision-making processes that Agentforce agents engage in throughout the customer journey, which includes activities like lead generation, opportunity qualification, and service case resolution, applicable across over 14 different sectors. Koa is specifically crafted for precise CRM functions and is assessed using Salesforce CRM Bench, incorporating genuine workflows such as opportunity updates, case routing, and follow-up scheduling. According to Salesforce, Koa achieves an 11% increase in accuracy for selecting appropriate actions and recalls customer context with a reliability that is 2.1 times greater than previous models. This innovative approach not only enhances efficiency but also significantly improves the overall customer experience across the board.

Leanstral 1.5 is a model licensed under Apache-2.0, designed for effective proof engineering in Lean 4, aimed at enhancing the capabilities and accessibility of formal verification. It boasts a total of 119 billion parameters, with 6 billion of them being active, marking a significant improvement in performance for tasks such as theorem proving, agent-based proof engineering, and the verification of practical code. The development of Leanstral 1.5 involved a comprehensive three-stage training process, which included mid-training, supervised fine-tuning, and reinforcement learning utilizing CISPO. In a multiturn environment, the model is tasked with receiving a theorem statement, submitting a proof, and refining its approach based on feedback from the Lean compiler until the proof is either successfully compiled or the available resources are depleted. In the code agent setting, Leanstral functions similarly to a developer navigating a raw filesystem, allowing it to edit files, execute bash commands, and interact with the Lean language server to monitor goals, errors, and type information in real time. This innovative approach not only streamlines the proof engineering process but also significantly enhances the user experience in formal verification tasks.

Beam represents Reflection’s inaugural open-weight model, characterized as a sparse Mixture-of-Experts framework featuring a staggering 501 billion parameters, of which 23 billion are actively utilized, specifically crafted for tasks involving coding, reasoning, and agentic functions. Its prowess is derived from extensive pretraining and reinforcement learning, having been developed using a vast dataset of 23.8 trillion diverse, high-quality tokens sourced from the web, public domains, and proprietary licensed materials. With a targeted emphasis on enhancing coding and agentic capabilities, Beam is engineered to provide competitive open-weight performance while ensuring efficient inference compute. This model adeptly handles a wide array of tasks, including intricate software engineering, terminal operations, STEM activities, web searches, tool utilization, and general knowledge inquiries. Reinforcement learning techniques have been employed to bolster its abilities in multi-step reasoning, tool application, and responsiveness to environmental feedback. Moreover, users have the flexibility to manage the balance between efficiency and performance through an adjustable reasoning effort parameter, thus tailoring the model's output to better suit their specific needs.

Microsoft Frontier Tuning enables businesses to tailor one or multiple of Microsoft’s leading MAI models to fit their specific operational requirements, allowing for training in a secure setting rather than depending on a standard AI model. The customization process begins by outlining the objectives and criteria for success, followed by integrating data, workflows, and insights gathered from Microsoft 365 and other sources. Continuous improvement is achieved through ongoing training and iterative refinement, with the model being deployed in platforms like Microsoft Foundry or Copilot, where it can enhance itself based on actual usage patterns. This innovative approach ensures that the models are well-versed in the organization’s terminology, context, processes, and expertise while maintaining strict privacy and security for all data within the client’s ecosystem. Additionally, Microsoft Frontier Tuning empowers teams with greater control over their models, minimizes the risks of vendor lock-in, and maximizes the return on investment by providing cutting-edge performance paired with exceptional token efficiency. As a result, organizations can expect to see enhanced operational effectiveness and a stronger alignment with their unique business strategies.

Gemini 3.5 Flash is Google’s high-performance multimodal AI model built to deliver frontier-level intelligence, fast execution speeds, and advanced agentic capabilities for coding, automation, and enterprise workflows. As the first release in the Gemini 3.5 series, the model is designed to help developers, businesses, and users execute complex long-horizon tasks through AI-powered reasoning, workflow orchestration, and intelligent automation. Gemini 3.5 Flash combines powerful coding performance, multimodal understanding, and real-time responsiveness while outperforming earlier Gemini models and competing frontier AI systems across several coding and reasoning benchmarks. The model is optimized for agentic workflows, allowing it to plan, execute, and manage multi-step tasks such as software development, infrastructure management, document preparation, and business process automation through the updated Antigravity harness. Gemini 3.5 Flash can also deploy collaborative subagents that work together under supervision to complete demanding workflows more efficiently and at lower operational cost. Beyond coding and automation, the platform generates richer graphics, dynamic web interfaces, interactive animations, and advanced multimodal experiences that support developers and enterprise users building AI-driven applications. Google has integrated Gemini 3.5 Flash across the Gemini app, AI Mode in Google Search, Google AI Studio, Android Studio, Gemini Enterprise Agent Platform, and enterprise AI services to expand access to advanced AI capabilities globally. The model also powers Gemini Spark, Google’s new personal AI agent designed to operate continuously and assist users with digital life management and automated task execution.

Claude Sonnet 4.5 represents Anthropic's latest advancement in AI, crafted to thrive in extended coding environments, complex workflows, and heavy computational tasks while prioritizing safety and alignment. It sets new benchmarks with its top-tier performance on the SWE-bench Verified benchmark for software engineering and excels in the OSWorld benchmark for computer usage, demonstrating an impressive capacity to maintain concentration for over 30 hours on intricate, multi-step assignments. Enhancements in tool management, memory capabilities, and context interpretation empower the model to engage in more advanced reasoning, leading to a better grasp of various fields, including finance, law, and STEM, as well as a deeper understanding of coding intricacies. The system incorporates features for context editing and memory management, facilitating prolonged dialogues or multi-agent collaborations, while it also permits code execution and the generation of files within Claude applications. Deployed at AI Safety Level 3 (ASL-3), Sonnet 4.5 is equipped with classifiers that guard against inputs or outputs related to hazardous domains and includes defenses against prompt injection, ensuring a more secure interaction. This model signifies a significant leap forward in the intelligent automation of complex tasks, aiming to reshape how users engage with AI technologies.

doteval serves as an AI-driven evaluation workspace that streamlines the development of effective evaluations, aligns LLM judges, and establishes reinforcement learning rewards, all integrated into one platform. This tool provides an experience similar to Cursor, allowing users to edit evaluations-as-code using a YAML schema, which makes it possible to version evaluations through various checkpoints, substitute manual tasks with AI-generated differences, and assess evaluation runs in tight execution loops to ensure alignment with proprietary datasets. Additionally, doteval enables the creation of detailed rubrics and aligned graders, promoting quick iterations and the generation of high-quality evaluation datasets. Users can make informed decisions regarding model updates or prompt enhancements, as well as export specifications for reinforcement learning training purposes. By drastically speeding up the evaluation and reward creation process by a factor of 10 to 100, doteval proves to be an essential resource for advanced AI teams working on intricate model tasks. In summary, doteval not only enhances efficiency but also empowers teams to achieve superior evaluation outcomes with ease.

Gemini Computer Use is an agentic computer interaction capability built into Gemini 3.5 Flash. It enables developers and enterprises to create AI agents that can work across browser, desktop, and mobile environments by seeing interfaces, reasoning through tasks, and taking action. The capability was previously offered through a standalone Gemini 2.5 computer use model, but is now natively integrated into Gemini 3.5 Flash. This gives developers access to stronger performance for agentic computer use tasks while also combining with Gemini’s existing strengths in function calling, Search grounding, Maps grounding, and built-in tools. Gemini Computer Use is designed for long-horizon automation, continuous software testing, enterprise knowledge work, and workflows that span multiple professional applications. Developers can start building with the feature through the Gemini API or Gemini Enterprise Agent Platform. Google also provides a demo environment through Browserbase for testing the capability. Safety controls include targeted adversarial training for live-environment risks, optional explicit user confirmation for sensitive or irreversible actions, and automatic task stopping when indirect prompt injection is identified. Gemini Computer Use helps organizations build practical AI agents that can complete complex digital tasks while supporting sandboxing, human review, and strict access controls.

Step 3.5 Flash is a cutting-edge open-source foundational language model designed for advanced reasoning and agent-like capabilities, optimized for efficiency; it utilizes a sparse Mixture of Experts (MoE) architecture that activates only approximately 11 billion of its nearly 196 billion parameters per token, ensuring high-density intelligence and quick responsiveness. The model features a 3-way Multi-Token Prediction (MTP-3) mechanism that allows it to generate hundreds of tokens per second, facilitating complex multi-step reasoning and task execution while efficiently managing long contexts through a hybrid sliding window attention method that minimizes computational demands across extensive datasets or codebases. Its performance on reasoning, coding, and agentic tasks is formidable, often matching or surpassing that of much larger proprietary models, and it incorporates a scalable reinforcement learning system that enables continuous self-enhancement. Moreover, this innovative approach positions Step 3.5 Flash as a significant player in the field of AI language models, showcasing its potential to revolutionize various applications.

DeepSWE is an innovative and fully open-source coding agent that utilizes the Qwen3-32B foundation model, trained solely through reinforcement learning (RL) without any supervised fine-tuning or reliance on proprietary model distillation. Created with rLLM, which is Agentica’s open-source RL framework for language-based agents, DeepSWE operates as a functional agent within a simulated development environment facilitated by the R2E-Gym framework. This allows it to leverage a variety of tools, including a file editor, search capabilities, shell execution, and submission features, enabling the agent to efficiently navigate codebases, modify multiple files, compile code, run tests, and iteratively create patches or complete complex engineering tasks. Beyond simple code generation, DeepSWE showcases advanced emergent behaviors; when faced with bugs or new feature requests, it thoughtfully reasons through edge cases, searches for existing tests within the codebase, suggests patches, develops additional tests to prevent regressions, and adapts its cognitive approach based on the task at hand. This flexibility and capability make DeepSWE a powerful tool in the realm of software development.

GLM-5.1 represents the latest advancement in Z.ai’s GLM series, crafted as a cutting-edge, agent-focused AI model tailored for coding, reasoning, and managing long-term workflows. This iteration builds upon the framework of GLM-5, which employs a Mixture-of-Experts (MoE) architecture to achieve high performance without incurring excessive inference expenses, aligning with a larger initiative towards open-weight models that are accessible to developers. A significant emphasis of GLM-5.1 is on fostering agentic behavior, allowing it to plan, execute, and refine multi-step tasks instead of merely reacting to isolated prompts. Its capabilities are specifically engineered to manage intricate workflows, such as debugging code, exploring repositories, and performing sequential operations while maintaining context over time. In comparison to its predecessors, GLM-5.1 enhances reliability during lengthy interactions, ensuring coherence throughout extended sessions and minimizing failures in multi-step reasoning processes. Overall, this model signifies a leap forward in AI development, particularly in its ability to support complex task management seamlessly.

GPT-5.2-Codex is a next-generation coding model created to support advanced, agent-driven software development. Built on the GPT-5.2 architecture, it is fine-tuned specifically for real-world engineering tasks. The model excels at working across large codebases while preserving context over long sessions. It handles complex refactors, migrations, and multi-step implementations more reliably than previous Codex models. GPT-5.2-Codex demonstrates top-tier performance in realistic terminal environments. Enhanced tool-calling and improved factual accuracy make it suitable for production workflows. The model is also significantly stronger in cybersecurity-related tasks. It can assist with vulnerability research and defensive security analysis. GPT-5.2-Codex includes safeguards designed to support responsible deployment. It represents a major advancement in professional-grade coding AI.

Lightning Rod is an innovative AI platform that streamlines the process of converting chaotic, unstructured real-world information into polished, production-ready datasets and specialized AI models without the need for manual labeling. This platform allows users to create high-quality, citable question-answer pairs derived from various sources, including news articles, financial documents, and internal records, effectively transforming raw historical data into organized datasets suitable for supervised fine-tuning or reinforcement learning applications. Utilizing an agent-driven workflow, users can articulate their objectives, and the system autonomously collects relevant sources, formulates questions, evaluates outcomes based on actual events, and incorporates contextual grounding before model training. A significant advancement of this platform is its “future-as-label” approach, which leverages real-world results as training signals, enabling AI systems to learn directly from authentic outcomes at scale rather than depending on synthetic or manually curated data. This capability not only enhances the accuracy of AI models but also improves their adaptability to dynamic real-world scenarios. With Lightning Rod, organizations can harness the power of their data more effectively than ever before.

Olmo 3 represents a comprehensive family of open models featuring variations with 7 billion and 32 billion parameters, offering exceptional capabilities in base performance, reasoning, instruction, and reinforcement learning, while also providing transparency throughout the model development process, which includes access to raw training datasets, intermediate checkpoints, training scripts, extended context support (with a window of 65,536 tokens), and provenance tools. The foundation of these models is built upon the Dolma 3 dataset, which comprises approximately 9 trillion tokens and utilizes a careful blend of web content, scientific papers, programming code, and lengthy documents; this thorough pre-training, mid-training, and long-context approach culminates in base models that undergo post-training enhancements through supervised fine-tuning, preference optimization, and reinforcement learning with accountable rewards, resulting in the creation of the Think and Instruct variants. Notably, the 32 billion Think model has been recognized as the most powerful fully open reasoning model to date, demonstrating performance that closely rivals that of proprietary counterparts in areas such as mathematics, programming, and intricate reasoning tasks, thereby marking a significant advancement in open model development. This innovation underscores the potential for open-source models to compete with traditional, closed systems in various complex applications.

Laguna XS.2 represents Poolside’s innovative open-weight coding model, distinguished as the lightest and quickest member of the Laguna series. This model features a total of 33 billion parameters in a Mixture of Experts setup, with 3 billion parameters activated, and has been meticulously trained in-house using 30 trillion tokens. As the latest generation model accessible to the public, it embodies a second-generation architecture and marks Poolside’s inaugural open-weight offering, drawing from insights gained during the training of Laguna M.1 with synthetic data and reinforcement learning techniques. Specifically designed to enhance agentic coding workflows, Laguna XS.2 excels in coding, acting, and rapidly iterating, particularly within Poolside’s coding agent environment. This model is particularly advantageous for developers and teams seeking a lightweight, efficient coding solution rather than a more cumbersome frontier system. Released under the permissive Apache 2.0 license, it empowers the community to assess, fine-tune, quantize, and build upon its weights, fostering a collaborative development atmosphere. In essence, Laguna XS.2 not only provides a robust platform for agentic coding but also encourages innovation and experimentation among its users.

Trinity Large Thinking is an innovative open-source reasoning model crafted by Arcee AI, tailored for intricate, multi-step problem solving and workflows involving autonomous agents that necessitate extended planning and the use of various tools. This model features a sparse Mixture-of-Experts architecture, boasting a remarkable total of around 400 billion parameters, with approximately 13 billion being active for each token, which enhances its efficiency while ensuring robust reasoning capabilities across a range of tasks, including mathematical calculations, code generation, and comprehensive analysis. A notable advancement in this model is its ability to perform extended chain-of-thought reasoning, which allows it to produce intermediate "thinking traces" prior to delivering final solutions, thereby boosting accuracy and reliability in complex situations. Furthermore, Trinity Large Thinking accommodates a substantial context window of up to 262K tokens, allowing it to effectively process lengthy documents, retain context during prolonged interactions, and function seamlessly in continuous agent loops. This model's design reflects a commitment to pushing the boundaries of what automated reasoning systems can achieve.

Ring is a sophisticated trillion-parameter thinking model created by Ant Group, specifically tailored for real-world Agent workflows. It employs a Mixture of Experts architecture similar to that of Ling, activating approximately 63 billion parameters during each inference, and is particularly geared towards tasks such as coding agents, utilizing tools, collaborating with multiple tools, engineering development, conducting research analysis, and executing long-term tasks. Instead of merely striving for "smarter" outcomes, Ring prioritizes the reliable completion of intricate tasks while maintaining a cost-effective approach, effectively balancing quality, speed, and efficiency in production settings. The latest iteration, Ring-2.6-1T, incorporates an adjustable Reasoning Effort mechanism that features high and xhigh reasoning intensity levels, which allocates an adaptive reasoning budget according to the complexity of the task at hand. The high mode is specifically optimized for high-frequency Agent workflows, resulting in lower token costs and quicker multi-step execution, while also facilitating multi-turn interactions, tool collaboration, and task decomposition. As a result, Ring demonstrates a significant advancement in enhancing the capabilities of agents in various operational contexts.

Odyssey-3 represents an advanced foundational world model and dynamic system that forecasts the movement and interaction of objects across space while tracking the evolution of various situations over time. Utilizing an autoregressive diffusion transformer, it captures representations related to physics, dynamics, and the principles of cause-and-effect from an extensive collection of visual data, enabling it to simulate environments effectively and develop strategies for managing physical systems. Additionally, Odyssey-3 is capable of creating immersive environments based on user prompts and can instantly predict how these environments shift in response to actions taken by individuals or agents, featuring both first-person and third-person navigation along with flexible camera controls. The extensive world knowledge it acquires can be tailored to various systems by retraining an action decoder or policy using paired data of observations and corresponding actions. Even with limited amounts of task-specific data, Odyssey-3 has demonstrated its utility in controlling robotic arms, powering humanoid robots, operating vehicles, and producing synchronized multi-sensory observations. Furthermore, it is capable of facilitating interactive experiences that enhance user engagement and understanding.