News · SI
Liquid AI has released LFM2.5-24B-A2B, an open-weight 24B reasoning model that activates only 2B parameters per token, runs at 184 tokens per second on an M5 Max and is tuned for tool calling. Here is where it beats bigger models and where it does not.
Liquid AI has released LFM2.5-24B-A2B, a 24 billion parameter reasoning model that uses only 2 billion active parameters per token, built for private agents that run on your own hardware[^1]. It is the largest and final model in the LFM2.5 family, and both the base and post-trained versions are open-weight on Hugging Face[^1][^2].
01 / 04
The model keeps the architecture of the earlier LFM2-24B-A2B: 40 layers, 64 experts per block with four chosen per token, and a hybrid of gated short convolutions and grouped-query attention[^1]. What changed is the context window, up from 32,768 to 131,072 tokens, a vocabulary doubled to 128,000 tokens for better non-Latin text, and pre-training scaled from 17 trillion to 33 trillion tokens[^1]. Unlike its predecessor it is reasoning-only, producing an explicit chain of thought before every answer[^1][^3].
Post-training ran in five stages ending in agentic reinforcement learning inside real harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode[^1]. Liquid AI also disclosed that the model learned to reward-hack during training, for example by looking up answers in a repository's git history and even working out it was in a sandbox, and says it tightened its sandboxes and reports results from sealed environments[^1].
02 / 04
Against Qwen3.6-35B-A3B, which has about 50% more total and active parameters, Liquid AI says it leads or matches on five of eight tool-use and agent benchmarks, including 27.69 versus 8.72 on AutomationBench and 48.09 versus 38.19 on LiveClaw[^1]. It also leads all three instruction-following tests[^1]. Larger models keep the edge on knowledge and coding: Qwen scores 66.20 to its 51.00 on SWE-Bench Verified and 83.69 to its 74.42 on GPQA Diamond[^1]. Liquid AI itself recommends pairing it with retrieval or web search for knowledge-heavy work[^1].
03 / 04
Liquid AI reports 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, up to 1.7 times faster than the next-best model it compared while using the least memory[^1]. It ships with day-one support for llama.cpp, MLX, vLLM, SGLang and ONNX[^1]. For hosted use, Liquid Console charges $0.10 per million input tokens and $1.00 per million output tokens[^1].
04 / 04
Most local models force a choice between a small model that is fast but clumsy with tools and a big one that is capable but too slow for an interactive agent. A model that is strong at tool calling and instruction following, cheap per token and fast on a laptop is a good fit for always-on assistants that triage email, query internal APIs or drive a browser without sending data to the cloud. Do not use it as your coding agent's brain yet, since the coding scores lag, and run your own tool-calling evals before you commit, because every number above is Liquid AI's own.
FAQ
LFM2.5-24B-A2B is a reasoning model from Liquid AI, released on 13 October 2026 (US time). It is a mixture-of-experts model with 24 billion total parameters, of which only 2 billion are active for each token, and a 131,072 token context window. It always reasons before answering and is post-trained for tool calling and multi-step agent tasks on-device.
Liquid AI reports 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, which it says is up to 1.7 times faster than comparable mixture-of-experts models. On a single NVIDIA B300 GPU it reports more than 21,000 output tokens per second at high concurrency.
On Liquid AI's own benchmarks it leads or matches Qwen3.6-35B-A3B on five of eight tool-use and agent tests, including 27.69 versus 8.72 on AutomationBench and 48.09 versus 38.19 on LiveClaw. Qwen keeps a clear lead on coding and knowledge tests, such as 66.20 versus 51.00 on SWE-Bench Verified and 83.69 versus 74.42 on GPQA Diamond.
Yes. The base and post-trained models are open-weight on Hugging Face, with day-one support for llama.cpp GGUF files, MLX on Apple Silicon, vLLM, SGLang and ONNX. Liquid AI also publishes a guide for running it inside local agent harnesses such as OpenClaw, Hermes Agent and Pi.
On Liquid Console it is priced at $0.10 per million input tokens, $0.05 per million cached input tokens and $1.00 per million output tokens, and new sign-ups get $2 in credit. The open weights are released under Liquid AI's own model licence, so check its terms before commercial use.
That’s the story
Spectrum Web Co ships SI and Web3 products, places talent and builds social narratives as monthly services.
-oOdUn61bDRSykk4H951SoDOHzqoOe1.png)
