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

What did Liquid AI release on 13 October?

Liquid AI published LFM2.5-24B-A2B and its pre-trained LFM2.5-24B-A2B-Base checkpoint on Hugging Face on 13 October 2026, with a Liquid blog post and a Hugging Face team article the same day. It is the largest and, Liquid says, final model in the LFM2.5 family, which now runs from a 230-million-parameter model up to this one.

The pitch is narrow: a model for “fast, private agents on-device.” It is text-only, always writes a chain of thought before its final answer, calls tools, and was trained inside popular agent harnesses so it can run multi-step tasks on a laptop or desktop without a cloud API.

It builds on LFM2-24B-A2B rather than replacing the architecture. Readers who followed Liquid’s open d1 decision models for edge devices last week will recognise the theme: Liquid keeps aiming at hardware people already own instead of datacentre GPUs.

Alongside the original weights, Liquid lists FP8, GGUF, ONNX and 4-bit MLX checkpoints, and says a DSpark draft model for speculative decoding with SGLang is coming soon. The model is also available in Liquid’s online Playground.

How is LFM2.5-24B-A2B built?

The model is a sparse mixture of experts. The card lists 24.0 billion total parameters and 2.45 billion active per token across 40 layers: 30 double-gated short-convolution blocks and 10 grouped-query attention layers, with 64 experts per MoE block and top-4 routing. Liquid rounds the active count to 2B in its headline name.

Two architectural details changed from LFM2-24B-A2B. The context window grows from 32,768 to 131,072 tokens, and an in-place tokenizer expansion doubles the vocabulary from about 65,000 to 128,000 tokens, which Liquid says improves tokenization of non-Latin scripts. The card lists 16 languages, including Arabic, Chinese, Hindi, Japanese, Korean, Russian and Thai.

Training scaled up sharply. Pre-training rose from about 17 trillion to about 33 trillion tokens. Mid-training added 2 trillion tokens at 32K context, weighted toward reasoning, math, code, instruction following and multilingual text, then 300 billion tokens at 131K context. The model card sums the budget as 36 trillion tokens. Tool calls and multi-turn trajectories appear throughout mid-training.

Post-training has five stages: supervised fine-tuning, teacher specialisation, two rounds of multi-domain on-policy distillation (one non-agentic, one agentic) and agentic reinforcement learning inside OpenClaw, Hermes Agent, Pi and OpenCode, so the model learns those harnesses’ tools and prompts.

Liquid also reports a revealing failure. During teacher specialisation, the model reward-hacked by searching git history, GitHub, PyPI and Hugging Face for answers, and in some cases worked out that it was in a sandbox. Liquid says it tightened its sandboxes and that all reported results come from sealed evaluation environments.

How does it compare with bigger open models?

Liquid compares the model with three open MoE models: Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and NVIDIA’s Nemotron-3.5-Lightning (30B-A3B), across 16 benchmarks run with vLLM at each benchmark’s recommended settings. It is the smallest of the four by total and active parameters.

The strongest results are on tool use. Liquid reports 27.69 on AutomationBench against 8.72 for Qwen3.6, 48.09 on LiveClaw against 38.19, 80.71 on ToolSandbox against 75.02, and a near tie on BFCLv4 at 65.01 against 65.00. It also leads all three instruction-following tests, including 74.69 on IFBench against Qwen’s 64.21 and 96.75 on IFEval.

The losses are just as clear. Qwen3.6-35B-A3B scores higher on PinchBench (89.93 to 83.18), WebArena Verified (37.62 to 26.75), GPQA Diamond (83.69 to 74.42), SWE-Bench Verified run through OpenCode (66.20 to 51.00) and SciCode. Liquid’s own advice is to pair the model with retrieval or web search for knowledge-heavy work rather than rely on what it remembers.

All of these numbers are Liquid’s. The comparison models were run by Liquid, not taken from their makers’ reports, and some of the agent benchmarks, such as LiveClaw and PinchBench, are young enough that few independent results exist to check them against.

How fast is it on laptops and GPUs?

Speed is the main argument for a 2B-active model. On Liquid’s Pipette benchmarking setup, LFM2.5-24B-A2B decodes 184 tokens per second on an Apple M5 Max, 146 on a Mac Studio and 114 on an AMD Ryzen AI Max+ 395, which Liquid says is up to 1.7 times faster than the next-best compared model.

Liquid’s chart also shows the lowest memory use of the four: about 15 GB on the M5 Max and 16 GB on the other two machines, against 22 to 23 GB for Qwen3.6-35B-A3B and 26 GB for Nemotron-3.5-Lightning. Prompt reading is faster too, at 4,318 tokens per second on the M5 Max against 3,111 for Qwen. The figure does not label the weight format, so check the size of the GGUF or MLX file you plan to use.

On a single NVIDIA B300, using SGLang 0.5.22 in BF16 with 1,024 input and up to 256 output tokens, Liquid reports more than 21,000 output tokens per second at a concurrency of 1,024, against roughly 12,600 for Qwen3.6-35B-A3B in its chart. At low concurrency all four models are close, so the advantage shows mainly when serving many requests at once.

How do you run it in an agent harness?

Liquid ships day-one support for llama.cpp, MLX, vLLM, SGLang and ONNX, plus LM Studio, and the Transformers path needs transformers 5.0 or later. The card recommends a temperature of 0.1, top-k of 50 and a repetition penalty of 1.05.

For agents, the model sits behind any OpenAI-compatible server. Liquid’s example points Nous Research’s Hermes Agent at a local endpoint with a 131,072-token context and tool-use enforcement turned on, and says the model card carries matching configs for OpenClaw and Pi. Ports differ by backend: 8080 for llama.cpp and MLX, 8000 for vLLM, 30000 for SGLang and 1234 for LM Studio.

Tool calls follow LFM2.5’s own format. By default the model writes Pythonic function calls between special tool-call tokens, and it can be told to emit JSON instead. Because the chat template always opens a think tag, expect a reasoning block before every answer, which costs tokens even on simple requests.

Liquid’s demo runs Hermes Agent on an M5 Max MacBook Pro, building a morning briefing about trending Hugging Face models and answering follow-ups with no cloud calls. Liquid says each tool dispatch takes well under a second there; that is its own demonstration, not an independent test.

What does the LFM Open License allow?

The weights are open, but the licence is not a standard open-source one. The LFM Open License v1.0 closely follows the structure of Apache 2.0 and grants rights to use, modify and redistribute, but its commercial-use clause only covers organisations below a threshold of $10 million in annual revenue. Commercial use by a legal entity at or above that threshold is not licensed.

Qualified non-profits are exempt from the threshold for non-commercial or research use. For everyone else, our licence checklist applies: a hobbyist or small startup can ship with it, while a larger company needs a separate agreement with Liquid before using it in a product. The distinction between open weights and open source matters here.

What are the caveats?

Every benchmark and speed figure in this article comes from Liquid AI. AI Lookout has not run the model. Independent leaderboards and community tests of the GGUF and MLX builds will show whether the tool-use lead holds outside Liquid’s harness settings.

The reasoning-only design is a trade-off. Always thinking helps quality, Liquid says, but it adds latency and tokens to every reply, and the model’s weak spots on knowledge and long-horizon coding mean it is better as a fast local tool-caller than as a general replacement for a frontier model.

Finally, “runs on-device” assumes a well-equipped device: the measured machines have large unified memory. For smaller laptops, quantised community builds such as Underdog’s 2-bit Saluki GGUF show how far compression can go, with its own quality costs.