Your central hub for AI news and updates on Open Source. We're tracking the latest articles, discussions, tools, and videos from the last 7 days.
Fyra's Brief
Kimi K2.6, an open-weights model from Moonshot AI, won a day of an AI Coding Contest, beating other top models in a real-time sliding-tile puzzle.
Moonshot AI's Kimi K2.6 win marks a closing of the capability gap between open-weights models and the frontier labs, a significant development in the field of AI.
Fyra's Brief
Qwen Team introduced Qwen-Scope, a collection of Sparse Autoencoders for the Qwen 3.5 family, enabling advanced model interpretability features.
Qwen-Scope provides a powerful tool for AI professionals to gain insight into and control their models, potentially revolutionizing the field of model interpretability.
Fyra's Brief
Hugging Face releases Mistral Medium 3.5, a dense 128B AI model capable of instruct, reasoning, and coding tasks in a single set of weights.
The release of Mistral Medium 3.5 marks a significant advancement in AI modeling capabilities, offering a unified solution for various tasks and applications.
Opne-source local AI that learns how you edit video Discussion | Link...
Check back soon for video content
Community talk
Microsoft Presents "TRELLIS.2": An Open-Source, 4b-Parameter, Image-To-3D Model Producing Up To 1536³ PBR Textured Assets, Built On Native 3D VAES With 16× Spatial Compression, Delivering Efficient, Scalable, High-Fidelity Asset Generation.
[RELEASE] - Finally, my first TTS model is out! 🎙️ Flare-TTS 28M
Qwen3.6-27B at 72 tok/s on RTX 3090 on Windows using native vLLM (no WSL, no Docker), portable launcher and installer
DeepSeek released 'Thinking-with-Visual-Primitives' framework
AMD Hipfire - a new inference engine optimized for AMD GPU's
Qwen3.6-27B vs 35B, I prefer 35B but more people here post about 27B...
I got tired of memory systems that break when you spin up new agents or fail to track sub-agent sessions properly.
Built an open-source tool to manage AI agent configs — 888 stars later, asking the AI community for feedback
We are finally there: Qwen3.6-27B + agentic search; 95.7% SimpleQA on a single 3090, fully local
Mistral Medium 3.5 128b ggufs are fixed
Qwen3.6-27B-NVFP4 - images
Been using Qwen-3.6-27B-q8_k_xl + VSCode + RTX 6000 Pro As Daily Driver
PFlash: 10x prefill speedup over llama.cpp at 128K on a RTX 3090
Qwen3.6-27B - Closed-loop SVG Images
Got DFlash speculative decoding working on Qwen3.5-35B-A3B with an RTX 2080 SUPER 8GB
Phosphene local video and audio generation for Apple Silicon open source (LTX 2.3) [P]
Follow-up: Qwen3.6-27B on 1× RTX 3090 — pushing to ~218K context + ~50–66 TPS, tool calls now stable (PN12 fix)
Mistral Medium Is On The Way
Local model on coding has reached a certain threshold to be feasible for real work
Found 48 Vulnerabilities in Open Source Projects During Live Testing with Claude Opus 4.6
GBNF grammar tweak for faster Qwen3.6 35B-A3B and Qwen3.6 27B
We built an open-source proxy that enforces LLM agent rules at the API layer - 700 GitHub stars
Caliber: open-source community registry for AI agent config files (CLAUDE.md, .cursor/rules, GEMINI.md) — 888 stars
I spent years building a 103B-token Usenet corpus (1980–2013) and finally documented it [P]
Qwen 3.6 27B vs Gemma 4 31B - making Packman game!
Open Models - April 2026 - One of the best months of all time for Local LLMs?
I open-sourced 59 Claude Skills covering the full website lifecycle (brand, design, content, SEO, dev, ops, growth)
Maybe the open-source race is splitting into different kinds of “useful intelligence” now
Toothcomb is an open-source tool for analysing and fact-checking speech in real time.
Project Aurelia — A 3-model architecture (80B + 13B + 9B) that physically reacts to my real-time heart rate via mmWave radar, spatial awareness via Lidar, and Vibration via Accelerometer. All on a Framework Desktop + eGPU
Deep Dive: Voicebox — The free, local-first ElevenLabs alternative that just hit 22K stars.
Oh Calude how can i trust you...
Opensource self-improving agents: How our agent performance increased autonomously by 40%
Image 2 model is pretty good at creating promotional product photos