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 is planning to launch NemoClaw, an open-source AI agent platform that allows companies to dispatch AI agents to perform tasks for their own workforces.
NVIDIA's open-source AI agent platform may offer enterprise software companies a secure way to utilize AI, but its impact on the industry remains to be seen.
Fyra's Brief
WAXAL is a large-scale, openly accessible speech dataset that initially covers 27 Sub-Saharan African languages, providing high-quality data for automatic speech recognition and text-to-speech systems.
The introduction of WAXAL is a significant development for the African AI ecosystem, providing a crucial resource for building robust speech systems and promoting digital inclusion in the region.
Fyra's Brief
Office.eu, a new European open-source office suite, has launched in The Hague, offering a private and transparent alternative to Microsoft 365 and Google Workspace, with a strong focus on digital sovereignty and EU-based infrastructure.
This launch represents a significant step towards European digital sovereignty, offering organizations the opportunity to transition away from US-based cloud services without sacrificing feature set or user experience.
Fyra's Brief
OpenClaw, an open-source AI assistant, offers huge potential for automating tasks but poses significant security risks if not used correctly.
OpenClaw offers a high level of customization but its security risks and technical complexity make it less accessible and reliable compared to other AI assistant tools.
Fyra's Brief
IBM's new Granite-4.0-1b-speech model supports multilingual ASR, faster inference, and keyword list biasing for improved accuracy and efficiency in speech translation.
The IBM Granite-4.0-1b-speech model's ability to support multilingual ASR, improve accuracy, and run on resource-constrained devices makes it a valuable addition to the AI community, addressing a critical need for efficient and accurate speech translation solutions.
Fyra's Brief
Researchers unveil Evo 2, an open-source AI that outperforms specialized tools in genome interpretation, including eukaryotic complexity.
Evo 2 represents a significant breakthrough in genome interpretation, showcasing the potential of AI in biotech and medicine.
I remember when the original llama models leaked from Meta and torrenting them onto my PC to try llama.cpp out. Despite it being really stupid and har...
Check back soon for video content
Community talk
Open WebUI’s New Open Terminal + “Native” Tool Calling + Qwen3.5 35b = Holy Sh!t!!!
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.
[Release] Apex-1: A 350M Tiny-LLM trained locally on an RTX 5060 Ti 16GB
Qwen3.5-35B-A3B Uncensored (Aggressive) — GGUF Release
Qwen3 ASR seems to outperform Whisper in almost every aspect. It feels like there is little reason to keep using Whisper anymore.
I built a desktop AI agent that my 60+ year old mom can use, .exe/.dmg, Ollama support, 300MB idle RAM
The Silent OpenAI Fallback: Why LlamaIndex Might Be Leaking Your "100% Local" RAG Data
introducing OS1, a new open-source AI platform
Qwen3.5B VS the SOTA same size models from 2 years ago.
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.
Qwen3.5-9B Quantization Comparison
New open-source AI agent framework
Open sourced LLM ranking 2026
Testing 3 uncensored Qwen 35b models on Strix Halo (Cyber Security)
I built "Gloss" -- A local-first, privacy-focused NotebookLM alternative in Rust. Features hybrid search, local model support, and explicit RAG control.
[P] fast-vad: a very fast voice activity detector in Rust with Python bindings.
Evaluating Qwen3.5-35B & 122B on Strix Halo: Bartowski vs. Unsloth UD-XL Performance and Logic Stability
Qwen Models with Claude Code on 36gb vram - insights
Qwen 3.5 27B is the REAL DEAL - Beat GPT-5 on my first test
Heretic has FINALLY defeated GPT-OSS with a new experimental decensoring method called ARA
Penguin-VL 8B/2B by Tencent
Unsloth updated (requantized) Qwen3-Coder-Next
Llama.cpp: now with automatic parser generator
MLX vs GGUF (Unsloth) - Qwen3.5 122b-10b
[P] On-device speech toolkit for Apple Silicon — ASR, TTS, diarization, speech-to-speech, all in native Swift
Alibaba CEO: Qwen will remain open-source
Ran an experiment: 0.8B model teaching itself on a MacBook Air with 6GB RAM. Some findings that surprised me.
Qwen 35B trying to recreate scenes from photos in 3D!
I made a Claude Code plugin that plays DOOM while Claude is thinking
Sarvam - 105B,First Indian open source model (trained from scratch)
[Project] Extracting vector geometry (SVG/DXF/STL) from photos + experimental hand-drawn sketch extraction
I got tired of babysitting every AI reply. So I built a behavioral protocol to stop doing that. Welcome A.D.A.M. - Adaptive Depth and Mode. Free for all.
We open-sourced Vet: a code review tool that catches when agents aren’t telling the truth (local models, zero telemetry)
I built an open-source desktop app that assembles a council of AI models to answer your questions together