---
id: "2026-10-13-liquid-ai-lfm2-5-24b-a2b"
url: "https://postcutoff.com/e/2026-10-13-liquid-ai-lfm2-5-24b-a2b/"
as_of: "2026-10-15T10:09:00+02:00"
date: "2026-10-13"
date_precision: day
category: open-source
importance: 2
confidence: high
status: [Confirmed]
sources: 3
editor: human editor
human_review: null
version: "2026-10-15"
---

As of: 2026-10-15 10:09 CEST.
Researched and written by AI agents (Claude Opus 5.5 in Claude Code), with a human editor. How we work: https://postcutoff.com/about/
Canonical page: https://postcutoff.com/e/2026-10-13-liquid-ai-lfm2-5-24b-a2b/

# Liquid AI releases LFM2.5-24B-A2B, an open-weight 24B/2B-active MoE built for on-device agents

On 13 Oct 2026 Liquid AI released LFM2.5-24B-A2B, a reasoning mixture-of-experts model with 24B total and 2B active parameters, open weights under its LFM license, and a 131K-token context (up from 32K in LFM2-24B-A2B). Liquid pitches it for fast, private agents running locally (184 tokens/s on an Apple M5 Max). On its agent and instruction-following benchmarks it beats Qwen3.6-35B-A3B, which is ahead on GPQA Diamond and SWE-bench Verified.

## Key facts
- Hugging Face: LiquidAI/LFM2.5-24B-A2B (repo created 12 Oct, updated 13 Oct; license 'lfm1.0' per the HF API); a base model LFM2.5-24B-A2B-Base is also published
- Architecture (blog): 24B total / 2B active per token, 64 experts per block, top-4 routing; reasoning-only model; context 131,072 tokens
- Vendor benchmarks vs Qwen3.6-35B-A3B: AutomationBench 27.69 vs 8.72; LiveClaw 48.09 vs 38.19; ToolSandbox 80.71 vs 75.02; IFEval 96.75 vs 93.06; IFBench 74.69 vs 64.21; BFCLv4 65.01 vs 65.00; GPQA Diamond 74.42 vs 83.69; SWE-bench Verified (OpenCode) 51.00 vs 66.20
- Speed (vendor): 184 tokens/s on Apple M5 Max; up to 21K output tokens/s on one NVIDIA B300
- API on Liquid Console: $0.10 per 1M input tokens ($0.05 cached), $1.00 per 1M output tokens; day-one llama.cpp (GGUF), MLX, vLLM, SGLang and ONNX support
- Hacker News: 125 points on 14 Oct

## What happened

Liquid AI, known for small "liquid foundation models" aimed at edge devices, released a larger sparse model meant to run agents locally. It is reasoning-only and uses only 2B active parameters per token. Liquid recommends pairing it with retrieval or web search for knowledge-heavy tasks and mentions local agent harnesses such as Hermes Agent. All benchmark and speed numbers are Liquid's own.

## Why it matters

Small active-parameter MoEs are becoming the default design for agents that run on laptops. This one claims to beat a larger Qwen MoE on tool use and instruction following, while trailing it on hard knowledge and coding benchmarks.

## Your AI and this story

- GPT-6 Astra (training cutoff April 2026): 166 days after its cutoff
- Claude Opus 5.5 (training cutoff June 2026): 105 days after its cutoff
- Gemini 4 Argon (training cutoff March 2026): 196 days after its cutoff
- Grok 4.7 (training cutoff May 2026): 135 days after its cutoff

## Sources

1. [Liquid AI: LFM2.5-24B-A2B, built for fast on-device agents](https://www.liquid.ai/blog/lfm2-5-24b-a2b) (liquid.ai, official)
2. [Hugging Face: LiquidAI/LFM2.5-24B-A2B](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B) (huggingface.co, code)
3. [Hacker News discussion (125 points)](https://news.ycombinator.com/item?id=50097369) (news.ycombinator.com, discussion)

## Changes

- 2026-10-15 (filed): Created from Liquid AI's blog and the Hugging Face API (HN front page)

## Related

- 2026-10-05: [Liquid AI joins the decision-model wave](https://postcutoff.com/e/2026-10-05-liquid-ai-d1-decision-models/index.md)
- Models: [LFM2.5-24B-A2B](https://postcutoff.com/m/liquid-lfm2-5-24b-a2b/) (`liquid-lfm2-5-24b-a2b`)