Liquid AI releases LFM2.5-24B-A2B: an agent model with 2B active parameters

Liquid AI has released LFM2.5-24B-A2B, an open reasoning model with 24B parameters, of which only 2B work on each token. The company is betting on agents that run around the clock on your own hardware rather than on demand. The model beats larger competitors on tool-use tasks, produces 184 tok/s on an Apple M5 Max and more than 21K output tokens per second on a single NVIDIA B300. The weights, including the base version, are already on Hugging Face.

Why 2B active parameters change the economics of an agent

An agent doesn't answer in a single reply. It thinks, calls a tool, reads the result, and thinks again. A single task consumes thousands of reasoning tokens, and the cost of those tokens decides whether you can keep an agent running permanently. With a MoE model that has 2B active parameters, each token costs about as much as on a small dense model, while the knowledge capacity is that of a 24B model.

The architecture is inherited from LFM2-24B-A2B: 40 layers, 64 experts in each MoE block with top-4 routing, and a hybrid of gated short convolutions and GQA. What changed is something else. Context grew from 32K to 131K tokens, and the vocabulary doubled from 65K to 128K to tokenize non-Latin languages better. The model is now reasoning-only: unlike its predecessor, it always produces an explicit chain of reasoning before the answer.

How it was trained: 33T tokens and five stages

Pretraining grew from ~17T to ~33T tokens. Mid-training follows: 2T tokens at 32K context, focused on reasoning, math, code, instruction following and multilingual data, and 300B tokens at 131K, focused on long context. Tool calls and multi-turn trajectories appear in both phases.

Post-training turns the base model into an agent in five stages:

• supervised fine-tuning (SFT);
• teacher specialization;
• multi-domain on-policy distillation (MOPD) on non-agentic tasks;
• MOPD on agentic tasks;
• agentic RL inside real harnesses: OpenClaw, Hermes Agent, Pi and OpenCode.

Training directly in the harnesses where the model will later run closes the usual gap between benchmark and practice.

Where the model wins and where it loses

Liquid AI compared the model with Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning (30B-A3B) on 16 benchmarks, with all runs going through vLLM. Key results for LFM2.5-24B-A2B (Qwen3.6-35B-A3B in parentheses):

• AutomationBench: 27.69 (8.72), more than three times the nearest competitor;
• BFCLv4: 65.01 (65.00);
• ToolSandbox: 80.71 (75.02);
• LiveClaw: 48.09 (38.19);
• τ³-Bench Banking: 8.04 (7.63);
• IFEval: 96.75 (93.06); IFBench: 74.69 (64.21); Multi-IF: 84.38 (81.53);
• MATH-500: 96.04 (96.73).

The competitors score higher on knowledge and code. GPQA Diamond: 74.42 against 83.69 for Qwen. SWE-Bench Verified (OpenCode): 51.00 against 66.20 for Qwen and 53.40 for Nemotron. PinchBench: 83.18 against 89.93. WebArena Verified: 26.75 against 37.62. SciCode: 32.71 against 37.35. On HLE the model scores 7.40, and on AA-Omniscience it scores -24.90. Liquid AI itself advises adding retrieval or web search for tasks that require knowledge.

Speed and how to run it yourself

On an Apple M5 Max the model decodes at 184 tok/s, and on a Ryzen AI Max+ 395 at 114 tok/s. That is up to 1.7x faster than the closest MoE in the comparison, with minimal memory use. On GPUs under high concurrency, a single B300 produces more than 21K output tokens per second, the best result in its class.

Day-one support covers llama.cpp, MLX, vLLM, SGLang and ONNX. Transformers requires version 5.0.0 or newer. The practical setup is to serve the model behind any OpenAI-compatible backend and point the harness at it. Ports depend on the server: 8080 for llama.cpp and MLX, 8000 for vLLM, 30000 for SGLang, 1234 for LM Studio. For Hermes Agent, a few hermes config set commands are enough, with base_url, context_length 131072 and tool_use_enforcement enabled.

The takeaway for developers: if your task comes down to tool calling, automation and strict instruction following, a local model with 2B active parameters already competes with larger MoE models, without sending data to the cloud. For complex coding and encyclopedic questions, it is still better to use a larger model or add external search. Both versions, the standard one and Base, are open for fine-tuning on your own tasks.