Liquid AI's LFM2.5-2.6B runs a full tool-calling agent on-device
2.6B open weights, 128K context, 220 tok/s under 2.5GB, free local inference — and Mistral's open Shieldstral lets you steer moderation in plain English.
Copy markdown
A capable agent that never leaves your device
Liquid AI's LFM2.5-2.6B is open-weight and runs tool-calling agents at 220 tok/s in under 2.5GB on an M5 Max (about 30 tok/s on a phone), scoring 56.9 on BFCLv4 tool use and 59.4 on LiveCodeBench — competitive with 4-10B models. Inference is free, data never leaves the box, and it drops into llama.cpp, MLX, vLLM, and OpenClaw/Hermes agent harnesses.
Steer your own safety filter in plain English
Mistral's Shieldstral is a 3B multimodal safety classifier under Apache 2.0 that takes your policy as a plain-language prompt at inference time — no retraining, no fixed categories. It runs on a single 16GB GPU, returns calibrated yes/no probabilities for text and images, and reportedly beats models up to 7x its size.
Build this weekend: run a 284B open model on your own box
GGUF quants of DeepSeek V4-Flash (284B MoE, 13B active, 1M context, MIT) have landed, so you can run it locally through llama.cpp — or call the official API beta at $0.28 per million output tokens.