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
Nvidia has released NemoClaw, an enterprise-grade AI agent platform with advanced security and privacy features, built on top of OpenClaw.
Nvidia's NemoClaw platform aims to provide a secure and customizable AI agent solution, addressing the growing need for enterprise-grade AI governance.
Fyra's Brief
NanoClaw, an AI agent platform, and Docker announce a partnership to enable isolated execution of NanoClaw agents within Docker containers, enhancing AI security.
This partnership is a significant step towards improving AI security, and its adoption may influence the development of more secure AI agent platforms.
Fyra's Brief
The OpenClaw AI agent has caused a frenzy in China, with companies and developers reaping benefits from users' enthusiasm for AI services, despite the AI agent's limitations and security risks.
The OpenClaw AI agent represents an important opportunity for Chinese tech companies and developers to capitalize on users' enthusiasm for AI services, but it also raises concerns about security risks and the need for technical proficiency.
Fyra's Brief
Leanstral, the first open-source code agent for Lean 4, has been released, offering efficient and trustworthy code generation capabilities through formal proof verification.
Leanstral's release marks a significant advancement in trustworthy code generation, enabling efficient and verifiable AI-assisted programming
Open Voice OS is an open-source, community-driven voice AI platform for building customizable voice assistants across embedded, single-board, and desk...
It's available [b8297](https://github.com/ggml-org/llama.cpp/releases/tag/b8297) onwards. Get latest llama.cpp version. >This adds support for NVIDIA...
Community talk
H Company just released Holotron-12B. Developed with NVIDIA, it's a high-throughput, open-source, multimodal model engineered specifically for the age of computer-use agents. (Performance on par with Holo2/Qwen but with 2x higher throughput)
Openrouter stealth model Hunter/Healer Alpha has been officially confirmed as MiMo, and a new model is coming.
Krasis LLM Runtime: 8.9x prefill / 4.7x decode vs llama.cpp — Qwen3.5-122B on a single 5090, minimal RAM
Introducing Unsloth Studio: A new open-source web UI to train and run LLMs
Running a 9B coding model at home and hitting 100% on HumanEval - how Agent Zero made it happen
I'm currently working on a pure sample generator for traditional music production. I'm getting high fidelity, tempo synced, musical outputs, with high timbre control. It will be optimized for sub 7 Gigs of VRAM for local inference. It will be released entirely free for all to use.
Gwen3.5-27b 8 bit vs 16 bit, 10 runs
Two weeks ago, I posted here to see if people would be interested in an open-source local AI 3D model generator
Built an open source tool that can find precise coordinates of any picture
Omnicoder-Claude-4.6-Opus-Uncensored-GGUF
LLMs forget instructions the same way ADHD brains do. I built scaffolding for both. Research + open source.
Mistral Small 4 is kind of awful with images
I was hyped for Nemotron 3 4B and it completely disappointed me compared to Qwen 3.5 4B
AI for investment research
Qwen3.5-9B on document benchmarks: where it beats frontier models and where it doesn't.
Qwen3.5-35B GGUF quants (16–22 GiB) - KLD + speed comparison
OmniCoder-9B best vibe coding model for 8 GB Card
Qwen 3.5 122b - a10b is kind of shocking
Nvidia updated the Nemotron Super 3 122B A12B license to remove the rug-pull clauses
Qwen3.5-27B performs almost on par with 397B and GPT-5 mini in the Game Agent Coding League
[P] I got tired of PyTorch Geometric OOMing my laptop, so I wrote a C++ zero-copy graph engine to bypass RAM entirely.
llama.cpp build b8338 adds OpenVINO backend + NPU support for prefill + kvcache
55 → 282 tok/s: How I got Qwen3.5-397B running at speed on 4x RTX PRO 6000 Blackwell
Qwen3 TTS in C++ with 1.7B support, speaker encoding extraction, and desktop UI
Qwen3.5 35b is sure one the best local model (pulling above its weight)
Meta just open-sourced everything and i feel like i'm the only one losing my mind about it
Thanks to the Intel team for OpenVINO backend in llama.cpp
built an open-source local-first control plane for coding agents
A side project we started in 2019 accidentally turned into an AIOS and an AI agent platform
2000 TPS with QWEN 3.5 27b on RTX-5090
Opus now supports 1 million contexts
Lemonade v10: Linux NPU support and chock full of multi-modal capabilities
Fine-tuned Qwen 3.5 2B to beat same-quant 4B, 9B, 27B, and 35B on a real dictation cleanup task, full pipeline, code, and eval (RTX 4080 Super, under £1 compute)
Running Qwen3.5-35B-A3B and Nemotron-3-Super-120B-A12B on a 5060ti and 1080ti with llama.cpp (Fully on GPU for Qwen; 64GB RAM needed for Nemotron)
Omnicoder-9b SLAPS in Opencode
OmniCoder-9B | 9B coding agent fine-tuned on 425K agentic trajectories
GATED_DELTA_NET for vulkan merged in llama.cpp
DoomVLM is now Open Source - VLM models playing Doom
I spent 8+ hours benchmarking every MoE backend for Qwen3.5-397B NVFP4 on 4x RTX PRO 6000 (SM120). Here's what I found.
New Model: LeVo 2 (SongGeneration 2), an open-source music foundation model
Introducing Unsloth Studio: an open-source web UI to run and train AI models
Built an open source LLM agent for personal finance
[P] mlx-tune – Fine-tune LLMs on Apple Silicon with MLX (SFT, DPO, GRPO, VLM)
RTCC — Dead-simple CLI for OpenVoice V2 (zero-shot voice cloning, fully local)
Agentic pipeline that builds complete Godot games from a text prompt
Caliber: open-source CLI to generate tailored Claude/Cursor configs & MCP recommendations
You guys gotta try OpenCode + OSS LLM
Unsloth will no longer be making TQ1_0 quants
[D] I built SuperML: A plugin that gives coding agents expert-level ML knowledge with agentic memory (60% improvement vs. Claude Code)
Real-time video captioning in the browser with LFM2-VL on WebGPU
Autonomous company frameworks are gaining traction