Open Source

Models and libraries you can run yourself.

Oct 25

Five Qwen3.8-27B models match Fable 5 Using GVS5H Harness

Five quantized Qwen 3.8-27B models achieve performance parity with Fable 5 when evaluated using the GVS5H harness, a benchmarking framework available on GitHub.

github.com/rzk/6 days ago

Oct 22

OpenShell – A safe, private runtime for fleets of autonomous AI agents

OpenShell is NVIDIA's runtime environment designed to safely execute multiple autonomous AI agents in distributed systems while maintaining privacy protections. The project enables secure orchestration of agent fleets through a dedicated runtime framework.

github.com/niemandhier/8 days ago

Hugging Face’s New AI Biology Model Lets You Chat With DNA

Hugging Face has released a new AI biology model that enables conversational interaction with DNA sequences. The model allows users to query and analyze genetic information through natural language, bridging computational biology and accessible AI interfaces.

wired.com/WIRED/9 days ago

WorldTUI: I trained a 360M model to pretend to be a terminal

WorldTUI is a 360M parameter model trained to simulate terminal behavior and output. The project demonstrates using smaller models for interface emulation tasks rather than general-purpose applications.

simedw.com/alphax314/9 days ago

Oct 21

Recursive launches Flint, a laptop coding model it says was largely built by AI

Recursive has released Flint, a free coding model designed to run on laptops, developed primarily through its own AI research system. The release serves as an early demonstration of self-improving AI capabilities.

techcrunch.com/Aditya Mehta/9 days ago

DeepSeek open-sources new tool to slash costs and ‘battle’ closed-source AI

DeepSeek has open-sourced a new tool designed to reduce AI inference costs and compete with closed-source AI systems. The release represents the company's strategy to make AI more accessible and cost-effective through open alternatives.

scmp.com/South China Morning Post/10 days ago

SC

Oct 20

DeepInfer: High-performance inference framework for DeepSeek models

DeepInfer is an inference framework optimized for running DeepSeek models with high performance. It provides infrastructure to deploy and execute these models efficiently.

github.com/ggcr/11 days ago

Oct 19

Timestral: Mistral's time-series foundation models

Mistral has released Timestral, a foundation model designed specifically for time-series forecasting tasks. The model builds on Mistral's existing capabilities to handle temporal data prediction.

mistral.ai/groar/12 days ago

Oct 18

Hybrid Reasoning Agents: LLM agents that think in latent space, act through text

Hybrid Reasoning Agents is a framework enabling LLM agents to perform reasoning in latent space while executing actions through text. The approach combines internal non-textual reasoning with external text-based interactions, available as an open-source implementation.

arxiv.org/allinonai/12 days ago

Show HN: Sphinx – A hook that quizzes you on your AI's code before shipping

Sphinx is a free and open-source hook that integrates with coding agent workflows to quiz developers on their understanding of code before production deployment. It automatically triggers questionnaires when code is pushed to production, ensuring developers comprehend the changes being shipped.

sphinx-six-theta.vercel.app/critiquebrowser/13 days ago

SP

NFS 2015 Offline Mode Arrives via Open - Source Emulator as EU Declines to Act

An open-source emulator enables offline play for NFS 2015, a game previously requiring always-online connectivity. The EU declined regulatory action on server shutdowns for online-dependent games.

techtimes.com/Richard L. Wells/13 days ago

Oct 17

Is the MCP protocol becoming the common foundation of smart , decentralised search ? · Contexte Tech

The Model Context Protocol (MCP) is emerging as a potential standardized foundation for building intelligent, decentralized search systems that operate across distributed networks.

contextetech.com/Tony, creator of Contexte Tech/14 days ago

DeepSeek-v4.1-Flash-1017

DeepSeek releases v4.1-Flash-1017, a lightweight language model variant optimized for faster inference and lower resource requirements. The model is available on Hugging Face and OpenRouter for immediate use.

huggingface.co/theykk/14 days ago

Oct 16

Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console

A developer has enabled Llama.cpp, a C++ implementation of the Llama language model, to run on a jailbroken PlayStation 5 console, allowing native execution of GGUF-format models on the hardware.

github.com/Anon84/14 days ago

Show HN: Gemini 4 Argon plays StarCraft

StarSkirmish Bench, a benchmark testing LLMs in StarCraft gameplay through code generation, has been updated with results from flagship models by OpenAI, Anthropic, Google, and Muse. The benchmark measures how well these models can handle complex strategic decision-making in a real-time strategy environment.

starskirmish.com/__cayenne/October 16, 2026

Oct 15

Meta launches an open-weight AI model that runs on Macs

Meta has released Muse Spark Open, an open-weight AI model designed to run locally on Apple silicon Macs through a partnership with Apple. The model enables on-device inference without requiring cloud infrastructure.

techcrunch.com/Lucas Ropek/October 15, 2026

Muse Spark Open

Meta releases Muse Spark Open, an open-source AI model available on Hugging Face. The model is released for public use and research purposes.

research.meta.ai/bigwheels/October 15, 2026

Oct 14

LFM2.5-24B-A2B: Built for fast on-device agents

Liquid AI releases LFM2.5-24B-A2B, a 24 billion parameter model optimized for fast, on-device agent execution. The model enables running autonomous agents directly on local hardware without cloud dependency.

liquid.ai/gmays/October 14, 2026

Oct 13

We turned EmbeddingGemma 2 into a speech-to-text model

A developer created a speech-to-text model by adapting EmbeddingGemma 2, Google's lightweight embedding model, for audio transcription tasks. The implementation is available for experimentation at the provided URL.

eg2.haku.sh/amrrs/October 13, 2026

Oct 12

Janas: Run LLMs larger than RAM by streaming experts from SSD

Janas is an open-source project that enables running large language models exceeding available RAM by streaming model experts from SSD storage. This approach allows users to work with larger models than their system memory would normally permit.

github.com/khalidkhair/October 12, 2026

QGIS bans AI assisted contributions

QGIS, the open-source geographic information system, has implemented a ban on contributions created or modified with AI assistance. The policy restricts code, documentation, and other submissions that involve AI tools in their creation or editing process.

github.com/bwnkl/October 12, 2026

Oct 10

Talorys – A self-hosted personal AI agent on Cloudflare's free tier

Talorys is a self-hosted personal AI agent that runs on Cloudflare's free tier, enabling users to deploy an autonomous AI assistant without incurring hosting costs. The project is available on GitHub for developers interested in running their own AI agents.

github.com/rociiu/October 10, 2026

ONLYOFFICE 10.0 brings AI tools, modular apps, and stronger security to its open-source office suite | Corporate

ONLYOFFICE 10.0 introduces AI-powered features, modular application architecture, and enhanced security measures to its open-source office productivity suite. The update expands functionality while maintaining the platform's self-hosted and open-source nature.

eqs-news.com/EQS News/October 10, 2026

EQ

Oct 9

Physics Meets AI: New Transfer Learning Tool Lets Weather Satellites Share Their Knowledge

A new transfer learning tool enables weather satellite data and models to be shared and applied across different satellites and observation scenarios. The approach combines physics-based constraints with machine learning to improve data efficiency and model generalization.

bioengineer.org/Bioengineer.org/October 9, 2026

Let your AI agents paint big arrows, boxes and text on your screen

A GitHub repository provides a tool enabling AI agents to draw arrows, boxes, and text directly onto screen output. The project appears to support visual annotations and UI elements for agent-based applications.

github.com/franze/October 9, 2026

Oct 7

AWS launches open-source AI agent sandbox to prevent YOLO mode disasters

AWS has released an open-source sandbox environment for AI agents designed to prevent uncontrolled autonomous agent behavior. The tool enables developers to test and contain agent actions before deployment in production systems.

theregister.com/The Register/October 7, 2026

AI-assisted proof of optimal packing for 11 squares

Researchers have formalized a computer-assisted proof demonstrating the optimal packing configuration for 11 squares within a unit square. The work combines automated reasoning with mathematical verification to solve a long-standing geometric problem.

github.com/bluepeter/October 7, 2026

Strands Decider 2B: a small, open-source, decision model

Strands introduces Decider 2B, a compact open-source model designed for decision-making tasks. The model appears positioned as a lightweight alternative for applications requiring autonomous decision capabilities.

strandsagents.com/gmays/October 7, 2026