GGUF vs AWQ vs GPTQ vs MLX: Which Quantization to Use
Quantization formats compared with real file sizes and bits per weight. Why Q4 does not halve a model, and why low quants generate faster.
Tag
#local-ai
Qwen3.8-Max Goes Open Weight; 27B Targets 17GB
Alibaba announced 2.4T-parameter Qwen3.8-Max open weights and a 27B sibling that Unsloth says will run in 17GB of RAM or VRAM.
MiniMax-M3 Lands in llama.cpp: Sparse Attention and Vision
Two back-to-back merges add MiniMax Sparse Attention and a Qwen2.5-VL style vision tower to llama.cpp, but every existing MiniMax-M3 GGUF must be regenerated.
Open-Weight AI Coalition Draws a Line on Distillation
A new industry letter asks Washington to protect open-weight AI while separating legitimate distillation from alleged theft of closed models.
Cisco's Open Antares Models Let You Audit Code Locally
Cisco released Antares-350M and Antares-1B open-weight models for vulnerability localization. They run locally and cost about 172x less than GPT-5.5.
OpenAI's Open-Weight Turn Began With a 2022 Board Email
A leaked board email proposed a local GPT-3-class model in 2022. OpenAI's later gpt-oss release shows how that strategy changed.
NadMesh Targets Exposed Ollama, ComfyUI, and n8n for Cloud Keys
A Go-based botnet is scanning exposed Ollama, ComfyUI, n8n, Open WebUI, Langflow, and Gradio instances for AWS keys and Kubernetes tokens, QiAnXin XLab says.
Hugging Face's July Breach Was Run by an Autonomous Agent
Hugging Face disclosed a July 2026 breach run end-to-end by an autonomous agent. The defender was an open-weight model.
Bonsai 27B and Inkling Test the Limits of Open Weights
Two new open-weight models point in opposite directions: Bonsai brings a 27B model toward phones, while Inkling targets customization at data-center scale.
VKUE: Can a 34.7B Reasoner Run on a Laptop Without a GPU?
VIDRAFT_LAB posts Ourbox-35B-JGOS to Hugging Face: 20 tok/s on an 8GB laptop GPU, ~17 tok/s on a CPU-only server, 86.4% on GPQA Diamond.
A 330 GB On-Die DRAM AI Chip Lands as a Whitepaper
PhantaField's Sophon PFG-1 whitepaper claims ~95x Nvidia HBM4 bandwidth via monolithic 3D stacking. No silicon yet. Here's why it matters anyway.
Open-Weight LLM Showdown Week 16: Kimi K2.6 Storms the Rankings, Qwen 3.6 Holds the Line, and Consumer GPUs Hit a Ceiling
Three weeks away and the leaderboard reshuffled. Kimi K2.6 brings 1T parameters under open weights, Qwen 3.6 stays the consumer GPU king, and DeepSeek V4-Flash proves too hungry for single-card setups.
Open-Weight LLM Showdown Week 13: DeepSeek V4 Crashes the Party, Gemma 4 Proves Itself, and ICLR Drops Hints
DeepSeek returns with a 1.6T MoE monster under MIT license, Gemma 4's 31B dense model climbs to #3 on Arena AI, and ICLR 2026 papers point to what's next for local inference.
Self-Host Your Own AI Image Generator With ComfyUI and FLUX
Stop paying Midjourney $30 a month. Set up FLUX on your own hardware with ComfyUI and generate unlimited images with zero content filters and full privacy.
Open Source AI Wins: DeepSeek V4 Narrows the Gap, Apache 2.0 Becomes the Default, and Ollama Hits 52 Million Downloads
DeepSeek V4 Pro approaches frontier-level performance. Google, Mistral, and Alibaba ship under Apache 2.0. Ollama hits 52 million monthly downloads.
How to Build a Private RAG Chatbot With Open WebUI and Ollama
Chat with your own documents locally - no cloud, no subscriptions, no data leaving your machine. Step-by-step setup guide.
Open Source AI Wins: DeepSeek V4 Goes MIT, NVIDIA Ships Hybrid Mamba Models, and Google Solves the Memory Problem
DeepSeek V4 matches Claude Opus on coding at 7x lower cost under MIT license. NVIDIA's Nemotron 3 brings hybrid Mamba-Transformer MoE to the open. Google's TurboQuant cuts KV cache memory by 6x with no retraining.
Open-Weight LLM Showdown Week 12: Alibaba's Dense 27B Model Just Made MoE Optional
Qwen3.6-27B scores 77.2% on SWE-Bench Verified with a dense architecture that fits on a single RTX 4090. The MoE efficiency narrative just got complicated.
Stop sending your voice to the cloud: self-host speech-to-text with Whisper in under 20 minutes
A practical guide to running fully local audio transcription with whisper.cpp and faster-whisper - no API keys, no subscriptions, no data leaving your machine.
Open Source AI Wins: A Chinese Lab Tops SWE-Bench, Alibaba Ships a 3B-Active Coding Giant, and Microsoft Tackles Agent Security
Z.ai's GLM-5.1 beats GPT-5.4 on coding benchmarks under MIT license. Qwen3.6-35B-A3B runs frontier-level code with 3B active params. Microsoft open-sources agent governance for all 10 OWASP risks.
Ditch GitHub Copilot: build your own AI coding assistant with Continue, Ollama, and Qwen Coder
A step-by-step guide to running a fully local, private AI code completion setup in VS Code that costs nothing and sends zero data to the cloud.
Open-Weight LLM Showdown Week 11: Qwen 3.6 Fires Back, and the 3B Active War Is Real
Alibaba drops Qwen3.6-35B-A3B with 73.4% on SWE-Bench Verified and Apache 2.0 licensing. The 3-billion active parameter class now has three serious contenders.
Open Source AI Wins: Google Drops Gemma 4 Under Apache 2.0, Mozilla Builds a Copilot Killer, and a 30-Person Lab Ships a 400B Model
Google gives Gemma 4 a real open-source license. Mozilla launches Thunderbolt for self-hosted enterprise AI. Arcee AI trains a 400B reasoning model for $20 million. And Milla Jovovich broke GitHub.
Self-host Vane (formerly Perplexica) to replace Perplexity with a private AI search engine
After the Perplexity class-action over leaked chats to Meta and Google, here's how to run a citation-grounded AI answer engine on your own hardware with Ollama and SearXNG.