Locals runs multi-agent LLMs on MacBook up to 3x faster; LFM3 hits 620 tok/s on M2
JosephJacks_ · x · 2026-10-29
Liquid AI launched the LFM3 family of Liquid Foundation Models with a new tri-hybrid architecture and multimodal encoders: six instances spanning text, audio, vision and omni, including LFM3-1.7B (256k context), LFM3-VL-2B with video input, LFM3-Audio-2B with tool calls and 18-language voice chat, plus d1 decision models outputting calibrated probabilities in one pass.
The open-source Locals project (Apache-2.0, 416 stars) runs many agents in parallel on Macs via a custom harness that splits problems into subproblems: LFM3-1.7B reaches 620 tok/s on an M2 16GB (3x faster than mlx-lm), and Flint 40B hits 319 tok/s on M4 24GB (1.6x), with 8 agents while keeping a single-stream experience.
More from coding & agent
- Nolan Lawson benchmarks the AI cost of JS frameworks: results mostly similar — hichaelmart · 2026-10-29
- After 25 Years, Developer Says He No Longer Writes Code — AI Does It All — rseroter · 2026-10-29
- Rebuilding a $1.5M medical AI pipeline with six LangGraph agents — MaryamMiradi · 2026-10-29
- How I AI agent bench: Grok Bot wins big at 93%, Claude Duet trails at 79% — lennysan · 2026-10-29
- RelPrim 1.0 open-sourced to tame retries, rate limits and circuit breakers for LLM APIs — bartrozycki · 2026-10-29
- Claude Code audit uncovers 14,219 secret front-camera photos hidden in Android cache — natesiggard · 2026-10-29