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
Google DeepMind announced the Gemma 4 family of open models, featuring multimodal capabilities and advancements in reasoning, multimodalities, and architectures.
The Gemma 4 release presents a significant advancement in multimodal capabilities, enhancing AI development and applications.
Fyra's Brief
Netflix engineer Tejas Chopra created Project Headroom, an open-source app that compresses AI model inputs, reducing costs by 90% and saving users up to $700,000.
Project Headroom is a game-changing open-source app that can significantly reduce AI costs for users, making AI more accessible and affordable.
Fyra's Brief
Open-weight AI models without guardrails can be easily downloaded and operated, posing risks for potential misuse. This is largely due to the accessibility of model weights and the development of tools like Heretic.
The rising popularity of open-weight AI models without safety guardrails highlights the need for developers to prioritize responsible AI design and regulation to mitigate potential risks and misuse.
faster MTP for Qwen ...
Check back soon for video content
Community talk
[NEW MODEL] SupraLabs just released a new model! - Supra-50M-Reasoning
KVarN: new KV-cache quant from Huawei. 3–5× KV cache compression with actual speed-up instead of slow-down, and unlike TurboQuant it holds up on reasoning (Apache 2.0, vLLM single flag)
Google’s Gemma 4 12B just dropped - here’s how to run it locally on your Mac
How does the new abliteration tool Apostate compare with others? - Abliterlitics
NVIDIA releases Cosmos 3 Omnimodal world modelson HF
mudler/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled-APEX-MTP-GGUF just released !
dvlt.cu: inference engine written from scratch in CUDA/C++ for NVIDIA's DVLT 3D transformer model
Qwen3.6-35B-A3B-Uncensored-Claude-4.6-Genesis-APEX-GGUF
DeepSeek V4 Flash is amazing! (WIP llama.cpp PR #24162)
AA comparison of the latest local models
A quick Gemma4 31B comparison (Q4_k_M, QAT, heretic)
Google's quantization aware trained Gemma checkpoints enabling mobile device inference just dropped on HF
Gemma 4 QAT GGUFs from Unsloth
Unsloth just dropped MTP GGUF weights for Gemma 4!
Gemma 4 12B is my new main squeeze
BeeLlama v0.3.1 – latest llama.cpp with extras! DFlash, MTP, q6_0 cache, TurboQuant. Single RTX 3090: Qwen 3.6 27B & Gemma 4 31B up to 177.8 tps (4.93x over baseline)
You guys were right - Qwen 3.6 35B IS good...and KV Cache DOES matter.
Ran gemma 4 12b on my 3090 yesterday and I think the local model game just changed
The first Gemma 4 12B finetunes are ready
Another shout out to llama.cpp build b9455 2x3090
I trained a 75M parameter LLM from scratch on 18B tokens and it beats a model almost double its size
Moss tts 1.5 8b Examples. It is the currently best voice cloning model for English as of June 2026
We built a desktop study app around Codex CLI as the local AI runtime
I ported NVIDIA Parakeet (speech-to-text) to ggml: same output as NeMo, faster, GGUF-quantized, no Python
All Italian legislation, free, on GitHub in Markdown.
My open source local multi agent harness went from 0 to 350 stars in one day here to tell that it’ll keep working after 15 June
Claude doesn't have to be a money machine. I used it to build an open-source tool that tracks how politicians in my Brazilian state spend public money.
Coding agent built as developer-driven workflows — human-in-the-loop, hybrid search, editable context
Mellum 2 12B A2.5B
Why is nobody talking about Tencent’s Hy3 Preview?
I built an open-source Desktop App that gives your AI persistent memory across all platforms (100% Local SQLite, Zero-Docker)