What is LFM2.5-24B-A2B?
LFM2.5-24B-A2B is the largest and last model in Liquid AI's LFM2.5 family, released on 13 October 2026. It is a mixture-of-experts model with 24 billion parameters, of which 2.45 billion work on each token, and an unusual layout: 30 of its 40 layers are short convolution blocks and only 10 use attention. It reads and writes text, has a 131,072-token context, and always reasons before answering. Compared with LFM2-24B-A2B it quadruples the context from 32K, doubles the vocabulary to 128,000 tokens and nearly doubles pre-training to about 33 trillion tokens.
Post-training included reinforcement learning inside four agent harnesses (OpenClaw, Hermes Agent, Pi and OpenCode), so the model has practised their tools and system prompts. Liquid says it decodes 184 tokens a second on an Apple M5 Max at 4-bit and reaches over 21,000 output tokens a second on one NVIDIA B300 at high concurrency.
Weights are under the LFM Open License v1.0, which bars commercial use by organisations with $10 million or more in annual revenue. Liquid's own API charges $0.10 per million input tokens and $1.00 per million output. Liquid has since released its next generation, LFM3, a family of small text, vision, audio, omni and decision models covered in our separate LFM3 review.
Is LFM2.5-24B-A2B worth it?
LFM2.5-24B-A2B is built for one job, calling tools on a local machine, and Liquid's numbers say it does that job better than larger rivals. Its card reports 27.69 on AutomationBench against 8.72 for Qwen3.6-35B-A3B, 48.09 on LiveClaw against 38.19, and 80.71 on ToolSandbox against 75.02, with all models run by Liquid in vLLM. It also leads all three instruction-following tests, with 96.75 on IFEval. Artificial Analysis had not scored the model when we checked, so none of this is independently confirmed.
The same table shows its ceiling. Qwen3.6-35B-A3B leads SWE-bench Verified in OpenCode at 66.20 to 51.00, GPQA Diamond at 83.69 to 74.42, WebArena Verified at 37.62 to 26.75 and PinchBench at 89.93 to 83.18. Liquid itself advises against using it alone for knowledge-heavy work and suggests pairing it with retrieval or web search.
The hardware case is strong. The Q4_K_M GGUF is 14.53 GB and a quantisation-aware Q4_0 build is 13.58 GB, so it fits a 24 GB Mac with room for context, or a 16 GB GPU with very little. Liquid measured 184 tokens a second on an M5 Max and 114 on an AMD Ryzen AI Max+ 395, up to 1.7 times the next-best model it compared, at a 4K-token input. Those are Liquid's figures from its own benchmarking rig.
The licence is the real constraint. The LFM Open License v1.0 grants commercial rights only to organisations under $10 million in annual revenue; anything larger needs a separate deal with Liquid. Gemma 4, under Apache 2.0, carries no such cap. Hosted use sidesteps the question, and OpenRouter's $0.05 input and $0.70 output per million undercuts Liquid's own $0.10 and $1.00, but every answer starts with a reasoning block, so output spend runs higher than a non-thinking model's.
Where it earns its place
A developer running Hermes Agent or OpenClaw on a 24 GB MacBook
Run a private personal agent that books, files and searches through tools
Serve the 14.53 GB Q4_K_M GGUF with llama.cpp or the MLX 4-bit build and point the harness at the local endpoint with a 131,072-token context. The model was trained in these harnesses; pair it with web search, since Liquid says knowledge is its weak side.
A platform engineer at a SaaS company serving automation agents to customers
Run high-volume tool-calling at low cost
Liquid reports over 21,000 output tokens a second on one B300 with SGLang at high concurrency. Self-hosting needs a commercial deal if revenue is $10 million or more; otherwise OpenRouter's $0.05/$0.70 rate puts 1 billion input and 100 million output tokens a month at about $120.
The founder of a seven-person legal-tech startup
Ship an on-device assistant inside a desktop product
Under $10 million in revenue the LFM licence allows commercial use and fine-tuning, with notices kept in redistributed copies. Budget for a renegotiation once the company crosses the threshold, because the licence then no longer covers it.
Before you commit
- Setup
- Transformers 5.0 or later, vLLM, SGLang, llama.cpp, MLX or LM Studio. A `<think>` tag is added to every reply by the chat template; Liquid recommends temperature 0.1, top_k 50 and repetition penalty 1.05.
- Integrations
- Tool calls default to Pythonic function lists between special tokens and switch to JSON if the system prompt asks; Liquid documents Hermes, OpenClaw and Pi configurations. Hosted on Liquid Console and on OpenRouter, where Liquid is the provider.
- Data and privacy
- Self-hosting keeps data local. The LFM Open License v1.0 allows commercial use only for organisations under $10 million in annual revenue, with a carve-out for non-profit research.
- Cost at scale
- On OpenRouter, 1 billion input and 100 million output tokens a month costs about $120; on Liquid Console's $0.10/$1.00, about $200. Self-hosting needs 13.58 to 14.53 GB at 4-bit, 25.49 GB at Q8_0 or 47.96 GB in BF16.
- Limits
- 131,072-token context; OpenRouter serves 128,000 tokens with an 8,192-token output cap. Text only. Always reasons first. Liquid advises against knowledge-heavy tasks without retrieval.
What's good
- 27.69 on AutomationBench vs 8.72 for Qwen3.6-35B-A3B (self-reported)
- 14.53 GB at Q4_K_M; 184 tokens/s on an M5 Max (Liquid)
- GGUF, MLX, ONNX and FP8 builds at launch
- $0.05 / $0.70 per 1M tokens on OpenRouter
What's not
- No commercial licence at $10M+ annual revenue
- Trails Qwen3.6-35B-A3B on SWE-bench Verified and GPQA
- Always reasons, so output tokens add up
- Text only; no independent leaderboard score yet
Best for
Developers and smaller companies running tool-calling agents on laptops or single GPUs, who fall under the $10 million revenue line or use a hosted API.
LFM2.5-24B-A2B alternatives
LFM2.5-24B-A2B: common questions
- Is LFM2.5-24B-A2B worth it?
- A laptop-sized agent model that out-calls Qwen3.6-35B-A3B on tool tests, held back by a licence that stops at $10 million in revenue. We rate it 3.8 out of 5.
- How much does LFM2.5-24B-A2B cost?
- Free weights under the LFM Open License (commercial use only below $10M annual revenue). Liquid Console: $0.10 input / $1.00 output per 1M tokens; OpenRouter: $0.05 / $0.70.
- What are the downsides of LFM2.5-24B-A2B?
- No commercial licence at $10M+ annual revenue; trails Qwen3.6-35B-A3B on SWE-bench Verified and GPQA; always reasons, so output tokens add up; text only; no independent leaderboard score yet.
- Who is LFM2.5-24B-A2B best for?
- Developers and smaller companies running tool-calling agents on laptops or single GPUs, who fall under the $10 million revenue line or use a hosted API.
- How much does LFM2.5-24B-A2B cost for a team or in production?
- On OpenRouter, 1 billion input and 100 million output tokens a month costs about $120; on Liquid Console's $0.10/$1.00, about $200. Self-hosting needs 13.58 to 14.53 GB at 4-bit, 25.49 GB at Q8_0 or 47.96 GB in BF16.
- Is LFM2.5-24B-A2B safe for company data?
- Self-hosting keeps data local. The LFM Open License v1.0 allows commercial use only for organisations under $10 million in annual revenue, with a carve-out for non-profit research.
- What are the best LFM2.5-24B-A2B alternatives?
- The closest alternatives we have reviewed are gpt-oss-20b, Gemma 4, Underdog Saluki 27B 1.0, Qwen3.8-27B and Ministral 3 (3B, 8B and 14B).
Answers checked Oct 28, 2026.