Liquid AI LFM2.5-24B-A2B: open-weight local agent model

Liquid AI ships open-weight 24B agent model for local use

Short answer

Liquid AI released LFM2.5-24B-A2B, an open-weight reasoning model with 24B parameters, of which 2B are active per token. It is built for tool-calling agents on local hardware and scores well on vendor-run tool-use and instruction-following benchmarks. It is weaker on knowledge-heavy and coding tests.

What this means for operators

If you run sales, support or operations at a 10-200 person company and want agents that call tools without sending every request to a hosted API, this is a model to test. Start with a narrow, rule-bound workflow such as ticket triage, CRM field updates or following a scripted process, where the reported strengths in tool use and instruction following apply. Do not give it knowledge-heavy questions without retrieval: Liquid AI itself recommends pairing the model with retrieval or web search for those. All scores come from the vendor's own evaluation, so run your own tasks before committing. The source does not state licence terms, so check the model card before commercial use.

Liquid AI has published LFM2.5-24B-A2B on Hugging Face, dated October 13, 2026. It is a reasoning model with 24 billion parameters in total and 2 billion active per token. It always produces a chain of thought before answering, calls tools and handles multi-step tasks. Both the model and a base version are open-weight and available for download.

What is in the release

The model keeps the architecture of its predecessor, LFM2-24B-A2B: 40 layers and 64 experts per mixture-of-experts block with top-4 routing. The context window grows from 32K to 131K tokens, and the vocabulary doubles from 65K to 128K tokens to handle non-Latin text better. Pre-training grew from about 17T to about 33T tokens. Post-training includes agentic reinforcement learning inside agent harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.

Reported performance

Liquid AI compared the model with Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning on 16 benchmarks, all run with vLLM. These are the vendor's own results.

AutomationBench: 27.69, against 8.72 for Qwen3.6-35B-A3B, the next-best model in the table.
LiveClaw: 48.09 against 38.19 for Qwen3.6-35B-A3B.
IFEval: 96.75 and IFBench: 74.69, ahead of the other three models in the table.

Larger models keep the lead elsewhere. On GPQA Diamond it scores 74.42 against 83.69 for Qwen3.6-35B-A3B. On SWE-Bench Verified with OpenCode it scores 51.00 against 66.20, and on WebArena Verified 26.75 against 37.62. Liquid AI advises pairing the model with retrieval or web search for knowledge-heavy work.

Speed and hardware

The source reports 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, up to 1.7x faster than the next-best compared model, with the lowest memory use. On a single NVIDIA B300 at high concurrency it reaches over 21K output tokens per second. Support at launch covers llama.cpp, MLX, vLLM, SGLang and ONNX.

How it plugs in

The model can be served locally behind any OpenAI-compatible backend, and an agent harness is then pointed at that endpoint. Liquid AI gives a Hermes Agent example with a 131072-token context setting and tool-use enforcement switched on. Fine-tuning tutorials are linked for adapting it to a specific task.

What the source does not say

The post gives no licence terms, no pricing and no results from production deployments. The benchmarks measure the model in a lab setting, not inside a company's own CRM or helpdesk.