Liquid AI's LFM3: six open 2B models you tune on one GPU
Text, vision, audio, omni and decision — all ~2B, all open-weight, all edge-ready. Plus a critical MCP Toolbox hole to patch before your agent leaks files.
Copy markdown
Six models, one 1.7B backbone
LFM3 is six open-weight ~2B models — text (256K context), vision with video, audio in/out across 18 languages, a vision+audio omni model, and two 'decision' models — all branching from one base, released Oct 27. The text model fits in ~1GB at 4-bit and decodes 1.8–2.8x faster on-device than the next-fastest rival.
Fine-tune your own on a single GPU
Every model is tunable on one GPU. Liquid's own demos: a wind-turbine defect spotter trained from ~3,000 photos in a few hours on one A100, and robot-arm orchestration via a rank-32 LoRA in ~6 GPU-hours on one MI325X. The unlock — ship a task-specific edge model this weekend instead of renting a frontier API.
Runs where you already deploy
Day-one support for llama.cpp, MLX, vLLM, SGLang and Qualcomm chips; weights are on Hugging Face, plus a hosted Playground and API. A full 128K context needs just 1.6GB of KV cache versus 4.3GB for Qwen3.5-4B, so it actually fits on the hardware you already have.
Patch MCP Toolbox before it leaks your files
CVE-2026-93328 (CVSS 9.3) in Google's MCP Toolbox for Databases lets an unauthenticated caller read host files, exfiltrate them to external Google Cloud Storage buckets, and overwrite local files via crafted tool parameters. It hits versions 1.2.0–1.4.0 — upgrade to 1.5.0+ and set allowedLocalRoots/allowedBuckets now.