Today's AI Frameworks & SDKs: Fastest-Growing Projects — September 21, 2026
Today's the AI Frameworks & SDKs space, there's a noticeable trend towards more efficient and innovative decision-making models, as well as multi-agent systems that enable autonomous exploration and improvement without requiring extensive training. These advancements aim to streamline processes and enhance performance across various applications.
wfzyx/von is an open-source System One decision model designed for sub-15ms, non-autoregressive decisions, offering a local drop-in alternative to TypeSafe Jev. With a strong growth score of 92.50 and 189 stars, it's clear that developers are attracted to its efficiency and performance in handling complex decision-making tasks locally.
yifanzhang-pro/recurrent-looped-tranformer (RLT) is the official project page for Recurrent Looped Transformer, which likely offers advanced capabilities in natural language processing or similar fields. Boasting 888 stars and a growth score of 71.67, it's evident that there's significant interest from developers looking to leverage its sophisticated transformer architecture.
AetherLabsAI/RSIAgent provides a training-free multi-agent framework for recursive self-improvement in new environments through broad-then-deep autonomous exploration and reusable memory. With 366 stars and a growth score of 44.44, this tool is gaining traction among researchers and developers interested in enabling agents to learn and adapt autonomously without the need for extensive training data.
mizorewww/laya-mlx offers a native MLX runtime tailored for Laya typed decision models, achieving short decisions at 7–14 ms on M3 Max. This tool has attracted considerable attention with 1,577 stars and a growth score of 35.10, highlighting its appeal to developers seeking efficient local decision-making capabilities without relying on text generation or cloud APIs.
0xBakeer/deepseek-v41-flash-spark integrates DeepSeek-V4.1-Flash technology with NVMe streaming and OpenAI API services, designed for optimal performance on a single DGX Spark (GB10) system. With 118 stars and a growth score of 34.27, it's growing due to its innovative approach to integrating multiple technologies to enhance decision-making processes.
Nehanth/swarmllm enables peer-to-peer LLM inference across browser tabs by splitting a large model over devices in a room through WebGPU and WebRTC runtime technology. This unique feature is driving interest with 425 stars and a growth score of 27.73, showcasing its potential for distributed computing and collaborative problem-solving.
hkqr/my-free-code serves as an open-source multi-provider AI gateway for Claude Code and other coding agents, facilitating model routing, streaming, tools, reasoning, fallbacks, and local model support. With 639 stars and a growth score of 27.72, it's growing due to its comprehensive approach to integrating multiple AI services into a cohesive development environment.
peonist-ai/halogen-flash-server provides the fastest way to run Qwen3.8-Flash-Next on Strix Halo (gfx1151), optimizing for performance and efficiency in specific hardware configurations. With 619 stars and a growth score of 23.04, it's growing due to its specialized focus on enabling optimal AI model execution in high-performance computing environments.
TokenRhythm/NeoHorse aims at recursive self-improvement via agentic post-training with routing harness, suggesting an innovative approach to enhancing the capabilities of machine learning models through autonomous improvement processes. With 691 stars and a growth score of 22.71, it's attracting developers interested in cutting-edge methodologies for model enhancement.
huawei-bayerlab/marigold-v2 revisits diffusion transformers for monocular depth estimation, offering advanced techniques for estimating depth from single images—a critical feature for applications like autonomous driving and robotics. With 779 stars and a growth score of 20.44, it's growing due to its specialized focus on improving the accuracy and efficiency of depth estimation in challenging environments.
Today's report highlights several tools that are making significant strides in enhancing decision-making capabilities, optimizing performance, and enabling autonomous learning within AI frameworks and SDKs.
wfzyx/von is an open-source System One decision model designed for sub-15ms, non-autoregressive decisions, offering a local drop-in alternative to TypeSafe Jev. With a strong growth score of 92.50 and 189 stars, it's clear that developers are attracted to its efficiency and performance in handling complex decision-making tasks locally.
yifanzhang-pro/recurrent-looped-tranformer (RLT) is the official project page for Recurrent Looped Transformer, which likely offers advanced capabilities in natural language processing or similar fields. Boasting 888 stars and a growth score of 71.67, it's evident that there's significant interest from developers looking to leverage its sophisticated transformer architecture.
AetherLabsAI/RSIAgent provides a training-free multi-agent framework for recursive self-improvement in new environments through broad-then-deep autonomous exploration and reusable memory. With 366 stars and a growth score of 44.44, this tool is gaining traction among researchers and developers interested in enabling agents to learn and adapt autonomously without the need for extensive training data.
mizorewww/laya-mlx offers a native MLX runtime tailored for Laya typed decision models, achieving short decisions at 7–14 ms on M3 Max. This tool has attracted considerable attention with 1,577 stars and a growth score of 35.10, highlighting its appeal to developers seeking efficient local decision-making capabilities without relying on text generation or cloud APIs.
0xBakeer/deepseek-v41-flash-spark integrates DeepSeek-V4.1-Flash technology with NVMe streaming and OpenAI API services, designed for optimal performance on a single DGX Spark (GB10) system. With 118 stars and a growth score of 34.27, it's growing due to its innovative approach to integrating multiple technologies to enhance decision-making processes.
Nehanth/swarmllm enables peer-to-peer LLM inference across browser tabs by splitting a large model over devices in a room through WebGPU and WebRTC runtime technology. This unique feature is driving interest with 425 stars and a growth score of 27.73, showcasing its potential for distributed computing and collaborative problem-solving.
hkqr/my-free-code serves as an open-source multi-provider AI gateway for Claude Code and other coding agents, facilitating model routing, streaming, tools, reasoning, fallbacks, and local model support. With 639 stars and a growth score of 27.72, it's growing due to its comprehensive approach to integrating multiple AI services into a cohesive development environment.
peonist-ai/halogen-flash-server provides the fastest way to run Qwen3.8-Flash-Next on Strix Halo (gfx1151), optimizing for performance and efficiency in specific hardware configurations. With 619 stars and a growth score of 23.04, it's growing due to its specialized focus on enabling optimal AI model execution in high-performance computing environments.
TokenRhythm/NeoHorse aims at recursive self-improvement via agentic post-training with routing harness, suggesting an innovative approach to enhancing the capabilities of machine learning models through autonomous improvement processes. With 691 stars and a growth score of 22.71, it's attracting developers interested in cutting-edge methodologies for model enhancement.
huawei-bayerlab/marigold-v2 revisits diffusion transformers for monocular depth estimation, offering advanced techniques for estimating depth from single images—a critical feature for applications like autonomous driving and robotics. With 779 stars and a growth score of 20.44, it's growing due to its specialized focus on improving the accuracy and efficiency of depth estimation in challenging environments.
Today's report highlights several tools that are making significant strides in enhancing decision-making capabilities, optimizing performance, and enabling autonomous learning within AI frameworks and SDKs.