Alternatives to BaseRT

Lucebox is a ready-to-use computer specifically designed for executing local AI models and agents at peak performance. Within its specially designed casing, it houses a Ryzen AI MAX+ 395 processor combined with 128GB of unified LPDDR5X memory and an RTX 3090 graphics card, both working in harmony through an open-source inference engine meticulously optimized for this configuration. The design of the architecture is key to its exceptional speed. The 128GB of unified memory allows large models to reside effectively, while the high-bandwidth VRAM of the 3090 serves as a rapid access tier. Techniques like speculative decoding (DFlash) and speculative prefill (PFlash) link these two memory systems, achieving inference speeds that can be up to 10 times faster than llama.cpp running on the same hardware, outperforming systems such as the Mac Studio and DGX Spark while being significantly more cost-effective. Moreover, this combination of hardware and software optimizations positions Lucebox as a formidable player in the local AI computing landscape.

LMCache is an innovative open-source Knowledge Delivery Network (KDN) that functions as a caching layer for serving large language models, enhancing inference speeds by allowing the reuse of key-value (KV) caches during repeated or overlapping calculations. This system facilitates rapid prompt caching, enabling LLMs to "prefill" recurring text just once, subsequently reusing those saved KV caches in various positions across different serving instances. By implementing this method, the time required to generate the first token is minimized, GPU cycles are conserved, and throughput is improved, particularly in contexts like multi-round question answering and retrieval-augmented generation. Additionally, LMCache offers features such as KV cache offloading, which allows caches to be moved from GPU to CPU or disk, enables cache sharing among instances, and supports disaggregated prefill to optimize resource efficiency. It works seamlessly with inference engines like vLLM and TGI, and is designed to accommodate compressed storage formats, blending techniques for cache merging, and a variety of backend storage solutions. Overall, the architecture of LMCache is geared toward maximizing performance and efficiency in language model inference applications.

LFM2 represents an advanced series of on-device foundation models designed to provide a remarkably swift generative-AI experience across a diverse array of devices. By utilizing a novel hybrid architecture, it achieves decoding and pre-filling speeds that are up to twice as fast as those of similar models, while also enhancing training efficiency by as much as three times compared to its predecessor. These models offer a perfect equilibrium of quality, latency, and memory utilization suitable for embedded system deployment, facilitating real-time, on-device AI functionality in smartphones, laptops, vehicles, wearables, and various other platforms, which results in millisecond inference, device durability, and complete data sovereignty. LFM2 is offered in three configurations featuring 0.35 billion, 0.7 billion, and 1.2 billion parameters, showcasing benchmark results that surpass similarly scaled models in areas including knowledge recall, mathematics, multilingual instruction adherence, and conversational dialogue assessments. With these capabilities, LFM2 not only enhances user experience but also sets a new standard for on-device AI performance.

EXAONE Deep represents a collection of advanced language models that are enhanced for reasoning, created by LG AI Research, and come in sizes of 2.4 billion, 7.8 billion, and 32 billion parameters. These models excel in a variety of reasoning challenges, particularly in areas such as mathematics and coding assessments. Significantly, the EXAONE Deep 2.4B model outshines other models of its size, while the 7.8B variant outperforms both open-weight models of similar dimensions and the proprietary reasoning model known as OpenAI o1-mini. Furthermore, the EXAONE Deep 32B model competes effectively with top-tier open-weight models in the field. The accompanying repository offers extensive documentation that includes performance assessments, quick-start guides for leveraging EXAONE Deep models with the Transformers library, detailed explanations of quantized EXAONE Deep weights formatted in AWQ and GGUF, as well as guidance on how to run these models locally through platforms like llama.cpp and Ollama. Additionally, this resource serves to enhance user understanding and accessibility to the capabilities of EXAONE Deep models.

Wafer is revolutionizing enterprise AI by offering the quickest open-source LLMs, enabling serverless and dedicated inference designed specifically for production workloads. With its serverless inference, teams can utilize top-tier open models without the burden of infrastructure and deployment challenges, providing rapid APIs that include GLM-5.2-Fast for reduced latency through EAGLE speculative decoding and a guaranteed throughput SLA, alongside GLM-5.2, which serves as a flagship model boasting enhanced coding and reasoning abilities. Wafer's innovative technology employs agents to optimize inference throughout the stack, pinpointing and addressing bottlenecks in orchestration, algorithms, serving engines, GPU kernels, and various hardware setups. This system meticulously profiles the stack to determine whether latency or throughput issues arise from factors such as scheduling, decoding, kernels, memory pressure, or hardware compatibility, and then it explores numerous paths to deliver the most effective solution. Rather than depending on a singular switch or heuristic, Wafer undertakes a comprehensive search of combinations involving models, engines, kernels, and hardware to maximize performance. By continually refining these combinations, Wafer ensures that enterprises can operate at peak efficiency while leveraging the best of open-source technologies.

vLLM is an advanced library tailored for the efficient inference and deployment of Large Language Models (LLMs). Initially created at the Sky Computing Lab at UC Berkeley, it has grown into a collaborative initiative enriched by contributions from both academic and industry sectors. The library excels in providing exceptional serving throughput by effectively handling attention key and value memory through its innovative PagedAttention mechanism. It accommodates continuous batching of incoming requests and employs optimized CUDA kernels, integrating technologies like FlashAttention and FlashInfer to significantly improve the speed of model execution. Furthermore, vLLM supports various quantization methods, including GPTQ, AWQ, INT4, INT8, and FP8, and incorporates speculative decoding features. Users enjoy a seamless experience by integrating easily with popular Hugging Face models and benefit from a variety of decoding algorithms, such as parallel sampling and beam search. Additionally, vLLM is designed to be compatible with a wide range of hardware, including NVIDIA GPUs, AMD CPUs and GPUs, and Intel CPUs, ensuring flexibility and accessibility for developers across different platforms. This broad compatibility makes vLLM a versatile choice for those looking to implement LLMs efficiently in diverse environments.

The Laguna XS 2.1 is an enhanced coding model that operates as an open weight agentic system, ideal for long-duration tasks on local machines. Featuring a 33-billion-parameter Mixture-of-Experts framework with 3 billion parameters activated per token, this model maintains the efficient architecture of Laguna XS.2 while significantly advancing performance in multilingual software engineering and terminal-style tasks. It is specifically engineered to assist coding agents in reviewing repositories, reasoning through intricate changes, utilizing various tools, executing commands, and maintaining continuity throughout extended projects. With a generous 256K context window, the model enables agents to effectively manage extensive codebases, lengthy histories, and complex multi-step workflows. Laguna XS 2.1 benefits from support from platforms like vLLM, SGLang, NVIDIA TensorRT-LLM, Hugging Face Transformers, and Ollama, with plans for native integration with llama.cpp in the future. The model is offered in various checkpoint formats, including BF16, FP8, INT4, and NVFP4, granting developers the flexibility to select between high fidelity and configurations optimized for limited VRAM or computational resources. This adaptability makes it an excellent choice for a wide range of development environments and requirements.

Mistral 7B is a language model with 7.3 billion parameters that demonstrates superior performance compared to larger models such as Llama 2 13B on a variety of benchmarks. It utilizes innovative techniques like Grouped-Query Attention (GQA) for improved inference speed and Sliding Window Attention (SWA) to manage lengthy sequences efficiently. Released under the Apache 2.0 license, Mistral 7B is readily available for deployment on different platforms, including both local setups and prominent cloud services. Furthermore, a specialized variant known as Mistral 7B Instruct has shown remarkable capabilities in following instructions, outperforming competitors like Llama 2 13B Chat in specific tasks. This versatility makes Mistral 7B an attractive option for developers and researchers alike.

Mirai is an advanced platform tailored for developers that focuses on on-device AI infrastructure, enabling the conversion, optimization, and execution of machine learning models directly on Apple devices with a strong emphasis on performance and user privacy. This platform offers a cohesive workflow that allows teams to efficiently convert and quantize models, assess their performance, distribute them, and conduct local inference seamlessly. Specifically designed for Apple Silicon, Mirai strives to achieve near-zero latency and zero inference cost, while ensuring that sensitive data processing remains securely on the user's device. Through its comprehensive SDK and inference engine, developers can swiftly integrate AI functionalities into their applications, leveraging hardware-aware optimizations to maximize the capabilities of the GPU and Neural Engine. Additionally, Mirai features dynamic routing abilities that intelligently determine the best execution path for requests, whether that be locally on the device or utilizing cloud resources, taking into account factors such as latency, privacy, and workload demands. This flexibility not only enhances the user experience but also allows developers to create more responsive and efficient applications tailored to their users' needs.

RunInfra effortlessly transforms natural language into fully operational AI inference endpoints. By simply describing your requirements, the AI agent autonomously constructs, refines, deploys, and scales your project without the need for YAML configurations, DevOps expertise, or GPU setup—just a conversation. Designed specifically for delivering open-source AI models as production-ready APIs, it intelligently chooses suitable models, benchmarks actual GPU performance, implements kernel enhancements, and establishes HTTP endpoints compatible with OpenAI. RunInfra is capable of creating diverse applications including language models, speech recognition, text-to-speech, embeddings, vision-language tasks, image generation, retrieval-augmented generation (RAG) searches, document analysis, transcription services, AI assistants, and complex multi-model reasoning frameworks, contingent on the runtime and model capabilities. Its streamlined workflow progresses seamlessly from your initial description to optimization, deployment, and integration; simply inform RunInfra of your needs, and it will evaluate real GPU options from L4 to B200, explore model variants like AWQ, GPTQ, and FP8, fine-tune kernels using Forge, and deliver a fully functional endpoint compatible with OpenAI’s Python and JavaScript SDKs. The efficiency and simplicity of RunInfra make it a valuable asset for developers aiming to leverage advanced AI technologies without the typical complexities involved.

NVIDIA TensorRT is a comprehensive suite of APIs designed for efficient deep learning inference, which includes a runtime for inference and model optimization tools that ensure minimal latency and maximum throughput in production scenarios. Leveraging the CUDA parallel programming architecture, TensorRT enhances neural network models from all leading frameworks, adjusting them for reduced precision while maintaining high accuracy, and facilitating their deployment across a variety of platforms including hyperscale data centers, workstations, laptops, and edge devices. It utilizes advanced techniques like quantization, fusion of layers and tensors, and precise kernel tuning applicable to all NVIDIA GPU types, ranging from edge devices to powerful data centers. Additionally, the TensorRT ecosystem features TensorRT-LLM, an open-source library designed to accelerate and refine the inference capabilities of contemporary large language models on the NVIDIA AI platform, allowing developers to test and modify new LLMs efficiently through a user-friendly Python API. This innovative approach not only enhances performance but also encourages rapid experimentation and adaptation in the evolving landscape of AI applications.

The Nexa SDK serves as a comprehensive developer toolkit that enables the local execution and deployment of any AI model on nearly any device equipped with NPUs, GPUs, and CPUs, facilitating smooth operation without reliance on cloud infrastructure. It features a rapid command-line interface, Python bindings, and mobile SDKs for both Android and iOS, along with compatibility for Linux, allowing developers to seamlessly incorporate AI capabilities into applications, IoT devices, automotive systems, and desktop environments with minimal setup and just one line of code to execute models. Additionally, it provides an OpenAI-compatible REST API and function calling, which simplifies the integration process with existing client systems. With its innovative NexaML inference engine, designed from the ground up to achieve optimal performance across all hardware configurations, the SDK accommodates various model formats such as GGUF, MLX, and its unique proprietary format. Comprehensive multimodal support is also included, catering to a wide range of tasks involving text, image, and audio, which encompasses functionalities like embeddings, reranking, speech recognition, and text-to-speech. Notably, the SDK emphasizes Day-0 support for the latest architectural advancements, ensuring developers can stay at the forefront of AI technology. This robust feature set positions Nexa SDK as a versatile and powerful tool for modern AI application development.

Qwen3.5-Plus is an advanced multimodal foundation model engineered to deliver efficient large-context reasoning across text, image, and video inputs. Powered by a hybrid architecture that merges linear attention mechanisms with a sparse mixture-of-experts framework, the model achieves state-of-the-art performance while reducing computational overhead. It supports deep thinking mode, enabling extended reasoning chains of up to 80K tokens and total context windows of up to 1 million tokens. Developers can leverage features such as structured output generation, function calling, web search, and integrated code interpretation to build intelligent agent workflows. The model is optimized for high throughput, supporting large token-per-minute limits and robust rate limits for enterprise-scale applications. Qwen3.5-Plus also includes explicit caching options to reduce costs during repeated inference tasks. With tiered pricing based on input and output tokens, organizations can scale usage predictably. OpenAI-compatible API endpoints make integration straightforward across existing AI stacks and developer tools. Designed for demanding applications, Qwen3.5-Plus excels in long-document analysis, multimodal reasoning, and advanced AI agent development.

Photon serves as the official high-performance inference engine for Moondream, specifically engineered to efficiently execute vision-language models across various platforms including cloud, desktop, and edge environments while ensuring real-time performance for AI applications in production. This advanced engine functions as a customized inference layer that is seamlessly integrated with the Moondream model framework, utilizing optimized scheduling, native image processing capabilities, and specialized CUDA kernels to enhance both speed and efficiency. Through this collaborative design, Photon achieves a remarkable reduction in latency compared to conventional vision-language model configurations, which facilitates quick interactions on edge devices and supports real-time data processing on server-grade systems. It boasts compatibility with a broad range of NVIDIA GPUs, accommodating everything from compact embedded systems like Jetson devices to powerful multi-GPU servers, thus providing versatility to meet varied operational demands. Additionally, Photon is equipped with production-ready features, including automatic batching, prefix caching, and memory-efficient attention mechanisms, further streamlining its performance in demanding scenarios. Such capabilities make it an ideal choice for developers seeking to implement AI-driven solutions across different environments.

LTM-2-mini operates with a context of 100 million tokens, which is comparable to around 10 million lines of code or roughly 750 novels. This model employs a sequence-dimension algorithm that is approximately 1000 times more cost-effective per decoded token than the attention mechanism used in Llama 3.1 405B when handling a 100 million token context window. Furthermore, the disparity in memory usage is significantly greater; utilizing Llama 3.1 405B with a 100 million token context necessitates 638 H100 GPUs per user solely for maintaining a single 100 million token key-value cache. Conversely, LTM-2-mini requires only a minuscule portion of a single H100's high-bandwidth memory for the same context, demonstrating its efficiency. This substantial difference makes LTM-2-mini an appealing option for applications needing extensive context processing without the hefty resource demands.

The Mixtral 8x7B model is an advanced sparse mixture of experts (SMoE) system that boasts open weights and is released under the Apache 2.0 license. This model demonstrates superior performance compared to Llama 2 70B across various benchmarks while achieving inference speeds that are six times faster. Recognized as the leading open-weight model with a flexible licensing framework, Mixtral also excels in terms of cost-efficiency and performance. Notably, it competes with and often surpasses GPT-3.5 in numerous established benchmarks, highlighting its significance in the field. Its combination of accessibility, speed, and effectiveness makes it a compelling choice for developers seeking high-performing AI solutions.

Fireworks collaborates with top generative AI researchers to provide the most efficient models at unparalleled speeds. It has been independently assessed and recognized as the fastest among all inference providers. You can leverage powerful models specifically selected by Fireworks, as well as our specialized multi-modal and function-calling models developed in-house. As the second most utilized open-source model provider, Fireworks impressively generates over a million images each day. Our API, which is compatible with OpenAI, simplifies the process of starting your projects with Fireworks. We ensure dedicated deployments for your models, guaranteeing both uptime and swift performance. Fireworks takes pride in its compliance with HIPAA and SOC2 standards while also providing secure VPC and VPN connectivity. You can meet your requirements for data privacy, as you retain ownership of your data and models. With Fireworks, serverless models are seamlessly hosted, eliminating the need for hardware configuration or model deployment. In addition to its rapid performance, Fireworks.ai is committed to enhancing your experience in serving generative AI models effectively. Ultimately, Fireworks stands out as a reliable partner for innovative AI solutions.

CodeQwen serves as the coding counterpart to Qwen, which is a series of large language models created by the Qwen team at Alibaba Cloud. Built on a transformer architecture that functions solely as a decoder, this model has undergone extensive pre-training using a vast dataset of code. It showcases robust code generation abilities and demonstrates impressive results across various benchmarking tests. With the capacity to comprehend and generate long contexts of up to 64,000 tokens, CodeQwen accommodates 92 programming languages and excels in tasks such as text-to-SQL queries and debugging. Engaging with CodeQwen is straightforward—you can initiate a conversation with just a few lines of code utilizing transformers. The foundation of this interaction relies on constructing the tokenizer and model using pre-existing methods, employing the generate function to facilitate dialogue guided by the chat template provided by the tokenizer. In alignment with our established practices, we implement the ChatML template tailored for chat models. This model adeptly completes code snippets based on the prompts it receives, delivering responses without the need for any further formatting adjustments, thereby enhancing the user experience. The seamless integration of these elements underscores the efficiency and versatility of CodeQwen in handling diverse coding tasks.

CodeT5 is an innovative pre-trained encoder-decoder model specifically designed for understanding and generating code. This model is identifier-aware and serves as a unified framework for various coding tasks. The official PyTorch implementation originates from a research paper presented at EMNLP 2021 by Salesforce Research. A notable variant, CodeT5-large-ntp-py, has been fine-tuned to excel in Python code generation, forming the core of our CodeRL approach and achieving groundbreaking results in the APPS Python competition-level program synthesis benchmark. This repository includes the necessary code for replicating the experiments conducted with CodeT5. Pre-trained on an extensive dataset of 8.35 million functions across eight programming languages—namely Python, Java, JavaScript, PHP, Ruby, Go, C, and C#—CodeT5 has demonstrated exceptional performance, attaining state-of-the-art results across 14 different sub-tasks in the code intelligence benchmark known as CodeXGLUE. Furthermore, it is capable of generating code directly from natural language descriptions, showcasing its versatility and effectiveness in coding applications.

Kluster.ai is an AI cloud platform tailored for developers, enabling quick deployment, scaling, and fine-tuning of large language models (LLMs) with remarkable efficiency. Crafted by developers with a focus on developer needs, it features Adaptive Inference, a versatile service that dynamically adjusts to varying workload demands, guaranteeing optimal processing performance and reliable turnaround times. This Adaptive Inference service includes three unique processing modes: real-time inference for tasks requiring minimal latency, asynchronous inference for budget-friendly management of tasks with flexible timing, and batch inference for the streamlined processing of large volumes of data. It accommodates an array of innovative multimodal models for various applications such as chat, vision, and coding, featuring models like Meta's Llama 4 Maverick and Scout, Qwen3-235B-A22B, DeepSeek-R1, and Gemma 3. Additionally, Kluster.ai provides an OpenAI-compatible API, simplifying the integration of these advanced models into developers' applications, and thereby enhancing their overall capabilities. This platform ultimately empowers developers to harness the full potential of AI technologies in their projects.

Falcon-40B is a causal decoder-only model consisting of 40 billion parameters, developed by TII and trained on 1 trillion tokens from RefinedWeb, supplemented with carefully selected datasets. It is distributed under the Apache 2.0 license. Why should you consider using Falcon-40B? This model stands out as the leading open-source option available, surpassing competitors like LLaMA, StableLM, RedPajama, and MPT, as evidenced by its ranking on the OpenLLM Leaderboard. Its design is specifically tailored for efficient inference, incorporating features such as FlashAttention and multiquery capabilities. Moreover, it is offered under a flexible Apache 2.0 license, permitting commercial applications without incurring royalties or facing restrictions. It's important to note that this is a raw, pretrained model and is generally recommended to be fine-tuned for optimal performance in most applications. If you need a version that is more adept at handling general instructions in a conversational format, you might want to explore Falcon-40B-Instruct as a potential alternative.

FriendliAI serves as an advanced generative AI infrastructure platform that delivers rapid, efficient, and dependable inference solutions tailored for production settings. The platform is equipped with an array of tools and services aimed at refining the deployment and operation of large language models (LLMs) alongside various generative AI tasks on a large scale. Among its key features is Friendli Endpoints, which empowers users to create and implement custom generative AI models, thereby reducing GPU expenses and hastening AI inference processes. Additionally, it facilitates smooth integration with well-known open-source models available on the Hugging Face Hub, ensuring exceptionally fast and high-performance inference capabilities. FriendliAI incorporates state-of-the-art technologies, including Iteration Batching, the Friendli DNN Library, Friendli TCache, and Native Quantization, all of which lead to impressive cost reductions (ranging from 50% to 90%), a significant decrease in GPU demands (up to 6 times fewer GPUs), enhanced throughput (up to 10.7 times), and a marked decrease in latency (up to 6.2 times). With its innovative approach, FriendliAI positions itself as a key player in the evolving landscape of generative AI solutions.

Liquid AI's LFM2.5 represents an advanced iteration of on-device AI foundation models, engineered to provide high-efficiency and performance for AI inference on edge devices like smartphones, laptops, vehicles, IoT systems, and embedded hardware without the need for cloud computing resources. This new version builds upon the earlier LFM2 framework by greatly enhancing the scale of pretraining and the stages of reinforcement learning, resulting in a suite of hybrid models that boast around 1.2 billion parameters while effectively balancing instruction adherence, reasoning skills, and multimodal functionalities for practical applications. The LFM2.5 series comprises various models including Base (for fine-tuning and personalization), Instruct (designed for general-purpose instruction), Japanese-optimized, Vision-Language, and Audio-Language variants, all meticulously crafted for rapid on-device inference even with stringent memory limitations. These models are also made available as open-weight options, facilitating deployment through platforms such as llama.cpp, MLX, vLLM, and ONNX, thus ensuring versatility for developers. With these enhancements, LFM2.5 positions itself as a robust solution for diverse AI-driven tasks in real-world environments.

Solar Mini is an advanced pre-trained large language model that matches the performance of GPT-3.5 while providing responses 2.5 times faster, all while maintaining a parameter count of under 30 billion. In December 2023, it secured the top position on the Hugging Face Open LLM Leaderboard by integrating a 32-layer Llama 2 framework, which was initialized with superior Mistral 7B weights, coupled with a novel method known as "depth up-scaling" (DUS) that enhances the model's depth efficiently without the need for intricate modules. Following the DUS implementation, the model undergoes further pretraining to restore and boost its performance, and it also includes instruction tuning in a question-and-answer format, particularly tailored for Korean, which sharpens its responsiveness to user prompts, while alignment tuning ensures its outputs align with human or sophisticated AI preferences. Solar Mini consistently surpasses rivals like Llama 2, Mistral 7B, Ko-Alpaca, and KULLM across a range of benchmarks, demonstrating that a smaller model can still deliver exceptional performance. This showcases the potential of innovative architectural strategies in the development of highly efficient AI models.

Pioneer serves as an inference API designed for developers who prioritize deployment over managing a GPU cluster. This tool allows teams to connect an existing client, such as OpenAI or Anthropic, to Pioneer, enabling them to maintain their API and code while performing inference seamlessly, all while Pioneer identifies areas where the current model may be lacking. It intelligently groups production traffic based on use cases, highlights opportunities for enhancement in accuracy, latency, or cost, and automatically creates and directs requests to specialized models. Through its continuous improvement mechanism known as Adaptive Inference, Pioneer analyzes real-time production failures to extract valuable examples, retrains a tailored model, assesses the updated checkpoint, and implements enhancements without necessitating any redeployment, all while maintaining access through the same endpoint. Additionally, Pioneer accommodates encoder models for tasks that require structured extraction, including named entity recognition, text classification, structured JSON extraction, privacy filtering, and safety classification, as well as decoder models that facilitate text generation, classification, and open-ended prompting. As a result, developers can optimize their workflows and enhance model performance with minimal hassle.

GigaChat 3 Ultra redefines open-source scale by delivering a 702B-parameter frontier model purpose-built for Russian and multilingual understanding. Designed with a modern MoE architecture, it achieves the reasoning strength of giant dense models while using only a fraction of active parameters per generation step. Its massive 14T-token training corpus includes natural human text, curated multilingual sources, extensive STEM materials, and billions of high-quality synthetic examples crafted to boost logic, math, and programming skills. This model is not a derivative or retrained foreign LLM—it is a ground-up build engineered to capture cultural nuance, linguistic accuracy, and reliable long-context performance. GigaChat 3 Ultra integrates seamlessly with open-source tooling like vLLM, sglang, DeepSeek-class architectures, and HuggingFace-based training stacks. It supports advanced capabilities including a code interpreter, improved chat template, memory system, contextual search reformulation, and 128K context windows. Benchmarking shows clear improvements over previous GigaChat generations and competitive results against global leaders in coding, reasoning, and cross-domain tasks. Overall, GigaChat 3 Ultra empowers teams to explore frontier-scale AI without sacrificing transparency, customizability, or ecosystem compatibility.

Tensormesh serves as an innovative caching layer designed for inference tasks involving large language models, allowing organizations to capitalize on intermediate computations, significantly minimize GPU consumption, and enhance both time-to-first-token and overall latency. By capturing and repurposing essential key-value cache states that would typically be discarded after each inference, it eliminates unnecessary computational efforts and achieves “up to 10x faster inference,” all while substantially reducing the strain on GPUs. The platform is versatile, accommodating both public cloud and on-premises deployments, and offers comprehensive observability, enterprise-level control, as well as SDKs/APIs and dashboards for seamless integration into existing inference frameworks, boasting compatibility with inference engines like vLLM right out of the box. Tensormesh prioritizes high performance at scale, enabling sub-millisecond repeated queries, and fine-tunes every aspect of inference from caching to computation, ensuring that organizations can maximize efficiency and responsiveness in their applications. In an increasingly competitive landscape, such enhancements provide a critical edge for companies aiming to leverage advanced language models effectively.

The LEAP Edge AI Platform presents a comprehensive on-device AI toolchain that allows developers to create edge AI applications, encompassing everything from model selection to inference directly on the device. This platform features a best-model search engine designed to identify the most suitable model based on specific tasks and device limitations, and it offers a collection of pre-trained model bundles that can be easily downloaded. Additionally, it provides fine-tuning resources, including GPU-optimized scripts, enabling customization of models like LFM2 for targeted applications. With support for vision-enabled functionalities across various platforms such as iOS, Android, and laptops, it also includes function-calling capabilities, allowing AI models to engage with external systems through structured outputs. For seamless deployment, LEAP offers an Edge SDK that empowers developers to load and query models locally, mimicking cloud API functionality while remaining completely offline, along with a model bundling service that facilitates the packaging of any compatible model or checkpoint into an optimized bundle for edge deployment. This comprehensive suite of tools ensures that developers have everything they need to build and deploy sophisticated AI applications efficiently and effectively.

This repository showcases the research preview of LongLLaMA, an advanced large language model that can manage extensive contexts of up to 256,000 tokens or potentially more. LongLLaMA is developed on the OpenLLaMA framework and has been fine-tuned utilizing the Focused Transformer (FoT) technique. The underlying code for LongLLaMA is derived from Code Llama. We are releasing a smaller 3B base variant of the LongLLaMA model, which is not instruction-tuned, under an open license (Apache 2.0), along with inference code that accommodates longer contexts available on Hugging Face. This model's weights can seamlessly replace LLaMA in existing systems designed for shorter contexts, specifically those handling up to 2048 tokens. Furthermore, we include evaluation results along with comparisons to the original OpenLLaMA models, thereby providing a comprehensive overview of LongLLaMA's capabilities in the realm of long-context processing.

Overshoot offers a comprehensive developer infrastructure for applications that require real-time video comprehension with minimal delay. Developers can initiate a stream and share inputs from cameras, screens, or videos via LiveKit, while also being able to query specific frames or time segments through hosted vision-language models via an OpenAI-compatible chat-completions interface. The platform efficiently manages the processes of live video ingestion, model serving, routing, stream lifecycle, and multimodal preprocessing. This allows applications to obtain either natural-language responses or structured JSON outputs without the need to maintain a dedicated GPU fleet or create a bespoke real-time video inference system. Overshoot is particularly well-suited for use cases in robotics, physical security, gaming, sports analytics, screen comprehension, industrial automation, and camera-based agents, making it a versatile tool for various industries. By simplifying the integration of live video processing, it empowers developers to focus on innovation rather than infrastructural complexities.

Xilinx's AI development platform for inference on its hardware includes a suite of optimized intellectual property (IP), tools, libraries, models, and example designs, all crafted to maximize efficiency and user-friendliness. This platform unlocks the capabilities of AI acceleration on Xilinx’s FPGAs and ACAPs, accommodating popular frameworks and the latest deep learning models for a wide array of tasks. It features an extensive collection of pre-optimized models that can be readily deployed on Xilinx devices, allowing users to quickly identify the most suitable model and initiate re-training for specific applications. Additionally, it offers a robust open-source quantizer that facilitates the quantization, calibration, and fine-tuning of both pruned and unpruned models. Users can also take advantage of the AI profiler, which performs a detailed layer-by-layer analysis to identify and resolve performance bottlenecks. Furthermore, the AI library provides open-source APIs in high-level C++ and Python, ensuring maximum portability across various environments, from edge devices to the cloud. Lastly, the efficient and scalable IP cores can be tailored to accommodate a diverse range of application requirements, making this platform a versatile solution for developers.

Luminal is a high-performance machine-learning framework designed with an emphasis on speed, simplicity, and composability, which utilizes static graphs and compiler-driven optimization to effectively manage complex neural networks. By transforming models into a set of minimal "primops"—comprising only 12 fundamental operations—Luminal can then implement compiler passes that swap these with optimized kernels tailored for specific devices, facilitating efficient execution across GPUs and other hardware. The framework incorporates modules, which serve as the foundational components of networks equipped with a standardized forward API, as well as the GraphTensor interface, allowing for typed tensors and graphs to be defined and executed at compile time. Maintaining a deliberately compact and modifiable core, Luminal encourages extensibility through the integration of external compilers that cater to various datatypes, devices, training methods, and quantization techniques. A quick-start guide is available to assist users in cloning the repository, constructing a simple "Hello World" model, or executing larger models like LLaMA 3 with GPU capabilities, thereby making it easier for developers to harness its potential. With its versatile design, Luminal stands out as a powerful tool for both novice and experienced practitioners in machine learning.

Yi-Large is an innovative proprietary large language model created by 01.AI, featuring an impressive context length of 32k and a cost structure of $2 for each million tokens for both inputs and outputs. Renowned for its superior natural language processing abilities, common-sense reasoning, and support for multiple languages, it competes effectively with top models such as GPT-4 and Claude3 across various evaluations. This model is particularly adept at handling tasks that involve intricate inference, accurate prediction, and comprehensive language comprehension, making it ideal for applications such as knowledge retrieval, data categorization, and the development of conversational chatbots that mimic human interaction. Built on a decoder-only transformer architecture, Yi-Large incorporates advanced features like pre-normalization and Group Query Attention, and it has been trained on an extensive, high-quality multilingual dataset to enhance its performance. The model's flexibility and economical pricing position it as a formidable player in the artificial intelligence landscape, especially for businesses looking to implement AI technologies on a global scale. Additionally, its ability to adapt to a wide range of use cases underscores its potential to revolutionize how organizations leverage language models for various needs.

Introducing the next iteration of our open-source large language model, this version features model weights along with initial code for the pretrained and fine-tuned Llama language models, which span from 7 billion to 70 billion parameters. The Llama 2 pretrained models have been developed using an impressive 2 trillion tokens and offer double the context length compared to their predecessor, Llama 1. Furthermore, the fine-tuned models have been enhanced through the analysis of over 1 million human annotations. Llama 2 demonstrates superior performance against various other open-source language models across multiple external benchmarks, excelling in areas such as reasoning, coding capabilities, proficiency, and knowledge assessments. For its training, Llama 2 utilized publicly accessible online data sources, while the fine-tuned variant, Llama-2-chat, incorporates publicly available instruction datasets along with the aforementioned extensive human annotations. Our initiative enjoys strong support from a diverse array of global stakeholders who are enthusiastic about our open approach to AI, including companies that have provided valuable early feedback and are eager to collaborate using Llama 2. The excitement surrounding Llama 2 signifies a pivotal shift in how AI can be developed and utilized collectively.

ChatGLM-6B is a bilingual dialogue model that supports both Chinese and English, built on the General Language Model (GLM) framework and features 6.2 billion parameters. Thanks to model quantization techniques, it can be easily run on standard consumer graphics cards, requiring only 6GB of video memory at the INT4 quantization level. This model employs methodologies akin to those found in ChatGPT but is specifically tailored to enhance Chinese question-and-answer interactions and dialogue. Following extensive training with approximately 1 trillion identifiers in both languages, along with additional supervision, fine-tuning, self-assistance through feedback, and reinforcement learning from human input, ChatGLM-6B has demonstrated an impressive capability to produce responses that resonate well with human users. Its adaptability and performance make it a valuable tool for bilingual communication.

Locai provides a robust infrastructure and tools designed to facilitate large-scale AI operations off the cloud. All processes occur on your own hardware, whether on-premises or within a private cloud, ensuring that your data remains within your control. The platform encompasses a comprehensive suite of functionalities. Locai Link serves as the inference runtime, enabling the deployment of large language models, audio transcription, video processing, and image recognition across various hardware types, including workstations, servers, and edge devices. Meanwhile, Locai Control functions as the orchestration layer, allowing users to oversee models, updates, and policies for entire fleets of machines from a single interface, ensuring seamless integration by enabling outputs from one model to feed into another for effective multimodal data fusion. To top it off, Locai Workspace offers teams a private AI environment equipped with chat, document management, and assistants, all supported by models that operate within the organization. Locai’s architecture is hardware-agnostic and compatible with a variety of models, including GGUF, ONNX, and TFLite. Developed in the UK with a focus on regulated sectors such as financial services, nuclear energy, defense, and critical infrastructure, Locai offers a free starting point for users to explore its capabilities. This adaptability makes Locai an appealing choice for organizations seeking secure AI solutions that fit their specific needs.

SiliconFlow is an advanced AI infrastructure platform tailored for developers, providing a comprehensive and scalable environment for executing, optimizing, and deploying both language and multimodal models. With its impressive speed, minimal latency, and high throughput, it ensures swift and dependable inference across various open-source and commercial models while offering versatile options such as serverless endpoints, dedicated computing resources, or private cloud solutions. The platform boasts a wide array of features, including integrated inference capabilities, fine-tuning pipelines, and guaranteed GPU access, all facilitated through an OpenAI-compatible API that comes equipped with built-in monitoring, observability, and intelligent scaling to optimize costs. For tasks that rely on diffusion, SiliconFlow includes the open-source OneDiff acceleration library, and its BizyAir runtime is designed to efficiently handle scalable multimodal workloads. Built with enterprise-level stability in mind, it incorporates essential features such as BYOC (Bring Your Own Cloud), strong security measures, and real-time performance metrics, making it an ideal choice for organizations looking to harness the power of AI effectively. Furthermore, SiliconFlow's user-friendly interface ensures that developers can easily navigate and leverage its capabilities to enhance their projects.

Our systems are globally implemented to ensure optimal response times for users everywhere. You can easily incorporate our OpenAI-compatible API into your application with minimal effort. Start the integration process in mere minutes and efficiently scale it to accommodate millions of users. Take advantage of our extensive scaling capabilities and performance enhancements, which allow our API to be four times more cost-effective than the OpenAI GPT-3.5 API. Experience the ability to generate text and engage in conversations with our AI model, which provides ChatGPT-level performance while being significantly more affordable. Getting started is a quick process, requiring only a few minutes with our API. Additionally, tap into the capabilities of one of the most advanced AI image models to produce breathtaking, high-quality images, graphics, and illustrations in just seconds, revolutionizing your creative projects. This approach not only streamlines your workflow but also enhances your overall productivity in content creation.

Effortlessly create, refine, and deploy high-performing, precise models using Deci’s deep learning development platform, which utilizes Neural Architecture Search. Achieve superior accuracy and runtime performance that surpass state-of-the-art models for any application and inference hardware in no time. Accelerate your path to production with automated tools, eliminating the need for endless iterations and a multitude of libraries. This platform empowers new applications on devices with limited resources or helps reduce cloud computing expenses by up to 80%. With Deci’s NAS-driven AutoNAC engine, you can automatically discover architectures that are both accurate and efficient, specifically tailored to your application, hardware, and performance goals. Additionally, streamline the process of compiling and quantizing your models with cutting-edge compilers while quickly assessing various production configurations. This innovative approach not only enhances productivity but also ensures that your models are optimized for any deployment scenario.

Sanctum serves as a private AI assistant that empowers users to operate and engage with comprehensive open-source LLMs directly on their devices. Constructed as a secure environment for AI, Sanctum ensures that all data remains encrypted and is confined to the user's computer. This platform simplifies the process of running AI locally, offering a user-friendly desktop application that enables instant setup of large language models on a Mac without the need for complex installations, and it operates entirely offline after the initial download. Prioritizing privacy, Sanctum features on-device processing and encryption, granting users full control over their data. With its integration with Hugging Face, users can effortlessly access a wide array of GGUF models, enabling them to verify compatibility, download models, and utilize them on either a PC or Mac. Additionally, Sanctum facilitates secure interactions with private PDF documents, allowing users to inquire, summarize, and engage with their files in a protected setting, thus enhancing the overall user experience. This level of accessibility and security positions Sanctum as a compelling choice for those seeking a personal AI solution that respects their privacy.

GLM-OCR is an advanced multimodal optical character recognition system and an open-source framework that excels in delivering precise, efficient, and thorough document comprehension by integrating textual and visual elements within a cohesive encoder-decoder design inspired by the GLM-V series. This model features a visual encoder that has been pre-trained on extensive image-text datasets alongside a streamlined cross-modal connector that channels information into a GLM-0.5B language decoder. It offers capabilities for layout detection, simultaneous recognition of various regions, and structured outputs for diverse content types, including text, tables, formulas, and intricate real-world document formats. Furthermore, it employs Multi-Token Prediction (MTP) loss and robust full-task reinforcement learning techniques to enhance training efficiency, boost recognition accuracy, and improve generalization across various tasks, leading to remarkable performance on significant document understanding challenges. This innovative approach not only sets new benchmarks but also opens up possibilities for further advancements in the field of document analysis.

On June 23, 2025, Microsoft unveiled Mu, an innovative 330-million-parameter encoder–decoder language model specifically crafted to enhance the agent experience within Windows environments by effectively translating natural language inquiries into function calls for Settings, all processed on-device via NPUs at a remarkable speed of over 100 tokens per second while ensuring impressive accuracy. By leveraging Phi Silica optimizations, Mu’s encoder–decoder design employs a fixed-length latent representation that significantly reduces both computational demands and memory usage, achieving a 47 percent reduction in first-token latency and a decoding speed that is 4.7 times greater on Qualcomm Hexagon NPUs when compared to other decoder-only models. Additionally, the model benefits from hardware-aware tuning techniques, which include a thoughtful 2/3–1/3 split of encoder and decoder parameters, shared weights for input and output embeddings, Dual LayerNorm, rotary positional embeddings, and grouped-query attention, allowing for swift inference rates exceeding 200 tokens per second on devices such as the Surface Laptop 7, along with sub-500 ms response times for settings-related queries. This combination of features positions Mu as a groundbreaking advancement in on-device language processing capabilities.

Oxlo.ai offers a privacy-centric inference platform tailored for agents, designed to operate cutting-edge open-source models while ensuring unlimited agentic tool utilization, secure failover, and complete absence of data retention or training. This platform provides developers with request-based access to a selection of curated open models via a streamlined HTTP API, which facilitates predictable usage, low-latency inference, and seamless integration into existing production environments. Teams can easily invoke models using OpenAI-compatible endpoints, transition from other service providers merely by adjusting the base URL and API key, and maintain support for a range of functionalities such as streaming, function calling, JSON mode, and various model types including vision models, embeddings, and image generation. With support for over 40 diverse models, Oxlo.ai encompasses a wide array of applications including text, chat, reasoning, coding, image generation, audio, embeddings, computer vision, vision-language, speech-to-text, text-to-speech, long-context, and detection workflows, making it a versatile tool for developers. This expansive support allows for innovative applications across multiple industries, enhancing the capabilities of teams looking to leverage advanced AI technologies.

LLaMA-Factory is an innovative open-source platform aimed at simplifying and improving the fine-tuning process for more than 100 Large Language Models (LLMs) and Vision-Language Models (VLMs). It accommodates a variety of fine-tuning methods such as Low-Rank Adaptation (LoRA), Quantized LoRA (QLoRA), and Prefix-Tuning, empowering users to personalize models with ease. The platform has shown remarkable performance enhancements; for example, its LoRA tuning achieves training speeds that are up to 3.7 times faster along with superior Rouge scores in advertising text generation tasks when compared to conventional techniques. Built with flexibility in mind, LLaMA-Factory's architecture supports an extensive array of model types and configurations. Users can seamlessly integrate their datasets and make use of the platform’s tools for optimized fine-tuning outcomes. Comprehensive documentation and a variety of examples are available to guide users through the fine-tuning process with confidence. Additionally, this platform encourages collaboration and sharing of techniques among the community, fostering an environment of continuous improvement and innovation.

Introducing DeepSeek-V3.2-Exp, our newest experimental model derived from V3.1-Terminus, featuring the innovative DeepSeek Sparse Attention (DSA) that enhances both training and inference speed for lengthy contexts. This DSA mechanism allows for precise sparse attention while maintaining output quality, leading to improved performance for tasks involving long contexts and a decrease in computational expenses. Benchmark tests reveal that V3.2-Exp matches the performance of V3.1-Terminus while achieving these efficiency improvements. The model is now fully operational across app, web, and API platforms. Additionally, to enhance accessibility, we have slashed DeepSeek API prices by over 50% effective immediately. During a transition period, users can still utilize V3.1-Terminus via a temporary API endpoint until October 15, 2025. DeepSeek encourages users to share their insights regarding DSA through our feedback portal. Complementing the launch, DeepSeek-V3.2-Exp has been made open-source, with model weights and essential technology—including crucial GPU kernels in TileLang and CUDA—accessible on Hugging Face. We look forward to seeing how the community engages with this advancement.