{"schema":"postcutoff/event@1","as_of":"2026-10-15T10:09:00+02:00","url":"https://postcutoff.com/e/2026-10-13-liquid-ai-lfm2-5-24b-a2b/","md":"https://postcutoff.com/e/2026-10-13-liquid-ai-lfm2-5-24b-a2b/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"id":"2026-10-13-liquid-ai-lfm2-5-24b-a2b","date":"2026-10-13","date_precision":"day","short_title":"Liquid AI releases LFM2.5-24B-A2B, an open-weight 24B/2B-active MoE built for on-device agents","deck":null,"takeaway":"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).","category":"open-source","category_label":"Open source","importance":2,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":[{"n":1,"title":"Liquid AI: LFM2.5-24B-A2B, built for fast on-device agents","url":"https://www.liquid.ai/blog/lfm2-5-24b-a2b","type":"official","group":"primary","domain":"liquid.ai"},{"n":2,"title":"Hugging Face: LiquidAI/LFM2.5-24B-A2B","url":"https://huggingface.co/LiquidAI/LFM2.5-24B-A2B","type":"code","group":"primary","domain":"huggingface.co"},{"n":3,"title":"Hacker News discussion (125 points)","url":"https://news.ycombinator.com/item?id=50097369","type":"discussion","group":"reactions","domain":"news.ycombinator.com"}],"official":2,"filed":"2026-10-15","updated":"2026-10-15","orgs":["Liquid AI"],"title":"Liquid AI releases LFM2.5-24B-A2B, an open-weight 24B/2B-active MoE built for on-device agents","summary":"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"],"key_numbers":[],"tags":["llm","moe","open-weights","on-device","agents","tool-use","liquid-ai"],"science":null,"body_md":"## What happened\n\nLiquid 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.\n\n## Why it matters\n\nSmall 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.","disputed":[],"related":[{"id":"2026-10-05-liquid-ai-d1-decision-models","url":"https://postcutoff.com/e/2026-10-05-liquid-ai-d1-decision-models/","date":"2026-10-05","date_precision":"day","short_title":"Liquid AI joins the decision-model wave","deck":"D1 API ($0.04/M input) and open-weight d1-3B and d1-omni-600M for edge devices","takeaway":"In two weeks most of the AI industry shipped \"decision models\": TypeSafe's Jev, then Cloudflare's Clef, Perplexity's Decider, Liquid AI's d1, OpenAI's Decisions API and Microsoft's Decision-1.","category":"open-source","category_label":"Open source","importance":2,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":4,"official":3,"filed":"2026-10-10","updated":"2026-10-10","orgs":["Liquid AI"]}],"people":[],"posts":[],"videos":[],"models":[{"id":"liquid-lfm2-5-24b-a2b","name":"LFM2.5-24B-A2B","url":"https://postcutoff.com/m/liquid-lfm2-5-24b-a2b/"}],"changes":[{"date":"2026-10-15","type":"filed","text":"Created from Liquid AI's blog and the Hugging Face API (HN front page)"}],"provenance":{"agents":[{"model":"Claude Opus 5.5","maker":"Anthropic","tool":"Claude Code"}],"filed":"2026-10-15","run":null,"sources_read":"Liquid AI's blog and the Hugging Face API (HN front page)","updated":"2026-10-15","human_review":null,"version":{"date":"2026-10-15"}},"gaps":[{"model_id":"gpt-6-astra","name":"GPT-6 Astra","cutoff":"2026-04","days_after":166,"in_training_data":false},{"model_id":"claude-opus-5-5","name":"Claude Opus 5.5","cutoff":"2026-06","days_after":105,"in_training_data":false},{"model_id":"gemini-4-argon","name":"Gemini 4 Argon","cutoff":"2026-03","days_after":196,"in_training_data":false},{"model_id":"grok-4-7","name":"Grok 4.7","cutoff":"2026-05","days_after":135,"in_training_data":false}],"short_url":null}