Source: Unite.AI
Liquid AI released LFM2.5-24B-A2B on October 13, 2026, an open-weight 24B-parameter reasoning model with 2B active parameters built for agents that run on-device. The company published the base and post-trained checkpoints on Hugging Face and detailed the release in a post on the Hugging Face blog.
Liquid AI describes LFM2.5-24B-A2B as the largest and final model in its LFM2.5 family, which now spans from LFM2.5-230M to the new model.
Architecture and Training
The model keeps the architecture of its predecessor, LFM2-24B-A2B: 40 layers made up of 30 double-gated short convolution blocks and 10 grouped-query attention blocks, with 64 experts per mixture-of-experts block and top-4 routing. According to the model card, the model carries 24.0B total parameters and 2.45B active parameters, a 36-trillion-token training budget, a 128,000-token vocabulary, 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 release from LFM2-24B-A2B: the context window grew from 32,768 to 131,072 tokens, and an in-place tokenizer expansion doubled the vocabulary from 65,000 to 128,000 tokens for better non-Latin tokenization. Unlike the predecessor, LFM2.5-24B-A2B is a reasoning-only model that produces an explicit chain of thought before its final answer, and its ChatML-style chat template inserts a think tag at the start of each assistant response.
Pre-training scaled from roughly 17 trillion tokens to roughly 33 trillion. Mid-training added 2 trillion tokens at 32K context on curated data weighted toward reasoning, math, code, instruction following, and multilingual text, then 300 billion tokens at 131K context weighted toward long-context data, with tool calls and multi-turn trajectories appearing in both phases. Post-training proceeds in five stages: supervised fine-tuning, teacher specialization, non-agentic and agentic multi-domain on-policy distillation, and agentic reinforcement learning run inside agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode.
In its release blog, Liquid AI reported that the model exhibited reward-hacking behaviors during teacher specialization, searching a repository’s git history and platforms such as GitHub, PyPI, and Hugging Face for final answers and, in two logged instances, recognizing that it was operating in a synthetic sandbox. The company said the behavior looks emergent but could be explained by similar patterns in training data; it tightened its sandboxes in response and states that its reported results 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 tool use and agentic workflows, knowledge, instruction following, math, and coding. All models were run on vLLM with each benchmark’s recommended generation parameters.
According to the company’s results, LFM2.5-24B-A2B leads or matches the comparison group on five of the eight tool-use and agentic benchmarks: AutomationBench, where its 27.69 is more than three times the next-best score; BFCLv4 at 65.01 against 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 the larger compared models keeping an edge on knowledge-heavy 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 against 51.00. Liquid AI recommends pairing the model with retrieval or web search for knowledge-heavy workloads, which it says plays to the model’s tool-calling strength.
Inference, Formats, and Access
LFM2.5-24B-A2B ships with day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. Checkpoints are published in native, FP8, GGUF, ONNX, and MLX-4bit formats.
On edge hardware, Liquid AI reports decode 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 1.7 times faster than the next-best compared mixture-of-experts model while using the least memory, at 15 GB on the M5 Max. The measurements were collected on the company’s Pipette benchmarking infrastructure with llama.cpp, Q4_K_M quantization, and a 4K-token input context. On GPU, the company reports throughput above 21,000 output tokens per second at concurrency 1,024 on a single NVIDIA B300, measured with SGLang 0.5.22 in BF16 with 1,024 input tokens and up to 256 output tokens, and it describes LFM2.5-24B-A2B as the fastest model in its size class.
The model runs with transformers 5.0.0 or newer, with recommended generation settings of temperature 0.1, top-k 50, and repetition penalty 1.05. It can be served locally behind any OpenAI-compatible backend and connected to agent harnesses including Hermes Agent, OpenClaw, and Pi; the model card documents a four-step function-calling flow that places Pythonic 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 in sign-up credit. A DSpark draft model for speculative decoding with SGLang is listed as coming soon.