A 24B model that thinks with only 2B at a time
On October 13, 2026, Liquid AI released LFM2.5-24B-A2B, the largest and final model in the LFM2.5 family (which spans from a 230M model up to this one). It is a mixture-of-experts reasoning model with 24 billion total parameters, but only about 2 billion activate per token — so each reasoning token costs about as much compute as a small 2B model. Because it is reasoning-only, every response begins with an explicit chain of thought before the final answer. Both the base checkpoint (LFM2.5-24B-A2B-Base) and the post-trained model are open-weight, available on Hugging Face and Liquid's Playground.
The spec sheet
- 24B total parameters, ~2B active per token (MoE, 64 experts per block, top-4 routing)
- 131,072-token context window (up from 32K)
- 128K-token vocabulary with in-place tokenizer expansion for non-Latin scripts
- Pre-trained on ~33T tokens, plus 2T-token mid-training
- Post-training: 5-stage pipeline (SFT, teacher specialization, two MOPD distillation rounds, agentic RL inside OpenClaw, Hermes Agent, Pi, OpenCode)
- Day-one support: llama.cpp, MLX, vLLM, SGLang, ONNX | API pricing: $0.10/1M input tokens, $1.00/1M output tokens
Benchmarks: beating bigger models at agent tasks
Liquid AI evaluated the model on 16 benchmarks with all models run on vLLM (these are the vendor's own results, not independently verified yet). It leads or matches the comparison group on five of eight tool-use and agentic benchmarks: AutomationBench 27.69 vs 8.72 for Qwen3.6-35B-A3B, LiveClaw 48.09 vs 38.19, BFCLv4 65.01 vs 65.00, ToolSandbox 80.71, and Tau3-Bench Banking 8.04. It also leads all three instruction-following benchmarks (IFEval 96.75, IFBench 74.69, Multi-IF 84.38). The larger compared models keep an edge on knowledge-heavy benchmarks like GPQA Diamond (74.42 vs 83.69) and coding benchmarks like SWE-Bench Verified (51.00 vs 66.20) — so pair it with retrieval or web search for knowledge-heavy work.
LFM2.5-24B-A2B vs the alternatives
| Model | Total params (active) | Context | Positioning |
|---|---|---|---|
| LFM2.5-24B-A2B (Liquid AI) | 24B (2B) | 131K | Open-weight reasoning model for on-device agents; 184 tok/s on M5 Max |
| Qwen3.6-35B-A3B (Alibaba) | 35B (3B) | long | Stronger on knowledge and coding benchmarks; larger footprint |
| Gemma-4-26B-A4B-it (Google) | 26B (4B) | long | Google ecosystem fit; trails on agentic benchmarks |
| Nemotron-3.5-Lightning (NVIDIA) | 30B (3B) | long | Inference-optimized; closest rival in size class |
How to use it: prompt it as an agent brain
LFM2.5-24B-A2B was post-trained inside agent harnesses (OpenClaw, Hermes Agent, Pi, OpenCode) with tool calls and multi-turn trajectories baked into mid-training. Prompt it with a goal, the tools it may call, and an acceptance criterion — then let its chain of thought plan the steps. Liquid AI even demonstrated it in Hermes Agent on an M5 Max MacBook Pro creating a morning briefing about trending Hugging Face models, entirely on-device with full audit trails and data never leaving the machine.
You are a local research agent running on my laptop. Goal: produce a morning briefing on this week's trending AI model releases. Steps: 1) query my saved list of AI news sources, 2) extract the 5 most important releases with dates and key specs, 3) for each, write a 2-line summary plus one line on who it matters for, 4) output as a clean briefing. Think step by step and call tools as needed. All data stays on this machine — nothing leaves it.
Frequently Asked Questions
Where can I download LFM2.5-24B-A2B?
Both the base (LFM2.5-24B-A2B-Base) and post-trained checkpoints are open-weight on Hugging Face, in Liquid AI's Playground, and supported day-one in llama.cpp (GGUF), MLX, vLLM, SGLang, and ONNX. For API inference, Liquid's console prices it at $0.10 per 1M input tokens and $1.00 per 1M output tokens.
What license does it use?
The model card lists the license as "other", named lfm1.0. Open-weight does not mean fully permissive open source — read Liquid AI's model license before commercial deployment.
Can it really run on a laptop?
Yes — that's the design point. Liquid AI reports 184 tokens/sec on an Apple M5 Max and 114 tokens/sec on an AMD Ryzen AI Max+ 395, with the least memory consumption among compared models. It needs transformers 5.0.0+ for the reference Python example.
Is it good at coding?
Moderately. It scores 51.00 on SWE-Bench Verified (via OpenCode) versus 66.20 for Qwen3.6-35B-A3B. Its strengths are tool use, instruction following, and agentic workflows — for heavy coding knowledge, pair it with a code-search tool.
Why does it always reason before answering?
It's a reasoning-only model by design: every response starts with an explicit chain of thought. Liquid AI adopted this to leverage the model's high decode speed (cheap reasoning tokens) to improve answer quality on multi-step tasks.
