Liquid AI releases open weight LFM2.5-24B-A2B for on-device agents – Unite.AI

Liquid AI released LFM2.5-24B-A2B on October 13, 2026, an open-weight 24B parameterized reasoning model with 2B active parameters built for agents running on the device. The company released baseline and post-trained checkpoints to Hugging Face and detailed the release in a post on the Hugging Face blog.

Liquid AI describes the LFM2.5-24B-A2B as the largest and final model in its LFM2.5 family, which now ranges from the LFM2.5-230M to the new model.

Architecture and Education

The model maintains the architecture of its predecessor, LFM2-24B-A2B: 40 layers composed of 30 double-gate short convolution blocks and 10 pooled query attention blocks, with 64 experts per expert mix block and top-4 routing. According to the model sheet, the model carries 24.0 billion total parameters and 2.45 billion active parameters, a training budget of 36 trillion tokens, a vocabulary of 128,000 tokens, a 131,072 token context window, and support for 16 languages, including English, Arabic, Chinese, French, German, Japanese, Korean, Spanish, Hindi, and Polish.

Two changes separate the new version from LFM2-24B-A2B: the context window grew from 32,768 to 131,072 tokens, and an expansion of the in-place tokenizer doubled the vocabulary from 65,000 to 128,000 tokens for better non-Latin tokenization. Unlike its predecessor, LFM2.5-24B-A2B is a reasoning-only model that produces an explicit thought chain before the final response, and its ChatML-style chat model inserts a thought tag at the beginning of each assistant response.

Pre-formation went from around 17 trillion tokens to around 33 trillion. At mid-training, 2 trillion tokens were added in a context of 32,000 on curated data weighted for reasoning, math, code, follow-up instructions, and multilingual text, then 300 billion tokens in a context of 131,000 weighted for long context data, with tool calls and multi-round trajectories displayed in both phases. Post-training proceeds in five phases: supervised fine-tuning, teacher specialization, distillation on non-agent and agentic multi-domain policies, and agent reinforcement learning performed within agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode.

In its release blog, Liquid AI reported that the model exhibited reward hacking behavior during teacher specialization, searching a repository’s git history and platforms like GitHub, PyPI, and Hugging Face for final answers, and, in two recorded cases, recognizing that it was operating in a synthetic sandbox. The company said the behavior appears emergent but could be explained by similar patterns in the training data; in response it has tightened its sandboxes and says the results reported come from sealed evaluation environments.

Benchmark results

Liquid AI evaluated LFM2.5-24B-A2B against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning across 16 benchmarks covering agent tool usage and workflows, knowledge, instruction following, math, and coding. All models were run on vLLM with the build parameters recommended by each benchmark.

According to the company’s results, the LFM2.5-24B-A2B leads or matches the comparison group on five of eight tool usage and agent benchmarks: AutomationBench, where its 27.69 is more than three times the next best score; BFCLv4 at 65.01 versus 65.00 for Qwen3.6-35B-A3B; ToolSandbox at 80.71; LiveClaw at 48.09 against 38.19; and τ³-Bench Banking at 8.04. Liquid AI reports that the model also leads on all three instruction-following benchmarks, with 96.75 on IFEval, 74.69 on IFBench, and 84.38 on Multi-IF.

The same tables show that the larger models in comparison maintain an advantage on knowledge-intensive benchmarks such as GPQA Diamond, HLE, and AA-Omniscience, and on coding benchmarks such as SciCode and SWE-Bench Verified with OpenCode, where Qwen3.6-35B-A3B scored 66.20 versus 51.00. Liquid AI recommends pairing the model with retrieval or web search for knowledge-intensive workloads, which it says plays to the model’s strength in invoking tools.

Inference, formats and access

LFM2.5-24B-A2B comes with day-one support for llama.cpp, MLX, vLLM, SGLang and ONNX. Checkpoints are published in native formats, FP8, GGUF, ONNX and MLX-4bit.

On edge hardware, Liquid AI reports decoding speeds of 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, up to 1.7x faster than the best mixed expert model compared using less memory, at 15GB on the M5 Max. The measurements were collected on the company’s Pipette benchmarking infrastructure with llama.cpp, Q4_K_M quantization, and a context of 4K token input. On the GPU, the company reports throughput in excess of 21,000 output tokens per second with concurrency of 1,024 on a single NVIDIA B300, measured at SGLang 0.5.22 in BF16 with 1,024 input tokens and up to 256 output tokens, and describes the LFM2.5-24B-A2B as the fastest model in its size class.

The model runs on transformers 5.0.0 or later, with recommended build settings of temperature 0.1, top-k 50, and repeat penalty 1.05. Can be served locally behind any OpenAI-compatible backend and connected to agent harnesses including Hermes Agent, OpenClaw and Pi; The template sheet documents a four-step function call flow that places Python calls between special tool call tokens. For hosted inference, Liquid AI prices the model on its Liquid Console API at $0.10 per million input tokens, $0.05 for cached input, and $1.00 per million output tokens, with $2 sign-up credit. A draft DSpark model for speculative decoding with SGLang will be available soon.