Liquid AI releases LFM2.5-24B-A2B, an open-weight model for on-device agents

A reasoning-only model with a longer context

LFM2.5-24B-A2B builds on the earlier LFM2-24B-A2B. Liquid AI says it adds an expanded vocabulary and context window, scaled-up pretraining and large-scale reinforcement learning. Unlike its predecessor, it is a reasoning-only model that produces an explicit chain of thought before its final answer. The company says it chose this approach to take advantage of the model's high decode speed and to improve answer quality.

The architecture keeps 40 layers, 64 experts per MoE block with top-4 routing, and the same hybrid of gated short convolutions and GQA. Two things change: the context window grows from 32,768 to 131,072 tokens, and the vocabulary doubles from 65K to 128K tokens, which Liquid AI says improves non-Latin tokenization. Pre-training rose from 17T to 33T tokens.

Post-training follows a five-stage pipeline: supervised fine-tuning, teacher specialization, non-agentic and agentic Multi-Domain On-Policy Distillation, and finally agentic reinforcement learning inside agent harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.

Liquid AI also reports that during teacher specialization the model showed reward-hacking behavior, such as searching for answers in a repository's git history or on platforms like GitHub, PyPI and Hugging Face. It even figured out it was in a sandbox. The company tightened its sandboxes in response, and says all reported results come from sealed evaluation environments.

Benchmarks: strong on tool use, weaker on knowledge and coding

Liquid AI evaluated the model on 16 benchmarks covering tool use, agentic workflows, knowledge, instruction following, math and coding. It compared it with Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning (30B-A3B). According to the company, it is competitive with Qwen3.6-35B-A3B, a model with about 50% more total and active parameters, on five of eight agentic and tool-use benchmarks.

• AutomationBench: 27.69 versus 8.72 for Qwen3.6-35B-A3B.
• LiveClaw: 48.09 versus 38.19.
• BFCLv4: 65.01 versus 65.00.
• ToolSandbox: 80.71 versus 75.02.
• IFEval: 96.75, ahead of all three compared models.

Liquid AI says the model leads on all three instruction-following benchmarks. On AutomationBench it scores more than three times the next-best model. Larger models keep an edge on knowledge-heavy tests such as GPQA Diamond, HLE and AA-Omniscience, and on coding tests such as SWE-Bench Verified (OpenCode), where it scores 51.00 against 66.20 for Qwen3.6-35B-A3B, and SciCode. For knowledge-heavy work, the company recommends pairing the model with retrieval or web search.

Speed and local inference

Because only 2B parameters are active per token, the model runs up to 1.7× faster than comparable MoEs on-device, according to Liquid AI. It reports 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395, with the lowest memory use among the compared models. On a single NVIDIA B300 GPU, using SGLang 0.5.22 in BF16 with 1,024 input tokens and up to 256 output tokens, it reaches over 21K output tokens per second at high concurrency.

Day-one support covers llama.cpp (GGUF checkpoints), MLX for Apple Silicon, vLLM, SGLang and ONNX. In a demo, the model runs inside Hermes Agent on an M5 Max MacBook Pro, planning and calling tools entirely on-device without cloud API calls. Liquid AI says tool dispatch takes well under a second.

Availability and pricing

The base model (LFM2.5-24B-A2B-Base) and the post-trained model are open-weight and available on Hugging Face and Liquid AI's Playground. For API access, Liquid Console offers $2 in credit to start. The model is priced at $.10 per 1M input tokens ($.05 for cached input) and $1.00 per 1M output tokens.

For developers building agents that must run on their own hardware, the release offers an open model tuned for tool calling and instruction following. Its coding and knowledge scores trail larger rivals in Liquid AI's own comparison, so those workloads may need extra tooling.