Liquid AI's LFM2.5-24B-A2B tested locally: 2B active params for on-device agents
JosephJacks_ · x · 2026-10-16
Fahd Mirza ran Liquid AI's new LFM2.5-24B-A2B locally: 24B total parameters with only 2B active per token, built for fast on-device agents.
- Served locally via llama.cpp (Q80) and plugged into the Hermes Agent framework
- The model autonomously found and fixed a hidden bug in a dairy factory dashboard, then verified the fix with curl
- Liquid's own charts put it ahead of Qwen3.6-35B-A3B on agent and tool use, behind on coding and web tasks
- Also stress-tested across 81 languages, deep-sea submarine physics and a tattoo simulation, with real strengths and real slips
Verdict: a promising on-device MoE agent model, though whether it truly beats Qwen depends on the full benchmark video.
More from coding & agent
- Dev mulls Rust rewrite of agent platform because speed-of-light compile times matter for agents — lucasmeijer · 2026-10-16
- Bland launches branded iMessage agents sharing personas with voice and SMS bots — ycombinator · 2026-10-16
- Codex drives a physical CAN-bus motor, eyeing a fully Codex-built robot — OpenAIDevs · 2026-10-16
- smallwords: MIT-licensed Python lib syncing controlled vocab across GBNF, JSON Schema, validators — Empty-Reflection4280 · 2026-10-16
- Fleet dev's own lesson: 200 prod-derived eval cases still meant always playing catch-up — BraceSproul · 2026-10-16
- Box AI Extract plus OpenAI Decisions API automates document routing — craigsdennis · 2026-10-16