LFM2.5-24B-A2B: Fast on-device reasoning agents

Liquid AI has released LFM2.5-24B-A2B, a new lightweight model built for running private AI agents directly on consumer devices without requiring cloud infrastructure. The model launched on October 13, 2026, and is immediately available as open-weight code on Hugging Face.

The core innovation is architectural efficiency. LFM2.5-24B-A2B uses a sparse mixture-of-experts design with 24 billion total parameters but only 2 billion active per token during inference. This sparse approach means each reasoning computation costs significantly less than dense models of comparable capability, enabling continuous agent operation on modest hardware.

What changed

Liquid AI announced LFM2.5-24B-A2B as an upgrade to its earlier LFM2-24B-A2B model. The new version expands the context window from 32,000 to 131,000 tokens and doubles the vocabulary from 65,000 to 128,000 tokens for improved handling of non-Latin scripts. The model is now explicitly optimized for reasoning: it produces a chain-of-thought before answering, making its logic transparent for agent systems that need to trace decision paths.

The training process reflects this agent focus. Liquid AI scaled pre-training from approximately 17 trillion to 33 trillion tokens. Mid-training added 2 trillion tokens at the standard 32,000 token context and 300 billion tokens at the expanded 131,000 token context, with emphasis on reasoning, mathematics, code, instruction-following, and multilingual material. Post-training involved five stages: supervised fine-tuning, teacher specialization, two rounds of multi-domain on-policy distillation, and agentic reinforcement learning across frameworks like OpenClaw, Hermes Agent, Pi, and OpenCode.

Both LFM2.5-24B-A2B (the trained agent version) and LFM2.5-24B-A2B-Base (the base model) are now available as open-weight releases.

Why it matters

The release addresses a real constraint in enterprise and consumer AI: running agents locally requires models small enough to fit on accessible hardware yet capable enough to handle multi-step reasoning and tool use reliably. Liquid AI builds personal AI with device-level context awareness, and this model extends that vision by enabling agents that keep all reasoning, data, and tool execution on your own machine.

Benchmark performance shows where this matters most. LFM2.5-24B-A2B leads on tool-use and agentic workflows: it scores 48.09 on LiveClaw versus 38.19 for Qwen3.6-35B-A3B, which has roughly 50% more total and active parameters. On AutomationBench it scores 27.69, more than three times higher than the next model. Instruction-following is another strength: the model leads across all three instruction-following benchmarks tested. On IFEval it scores 96.75, IFBench 74.69, and Multi-IF 84.38.

For knowledge-heavy tasks, larger models retain an advantage. On GPQA Diamond and other knowledge benchmarks, models with more parameters perform better. This means practitioners using LFM2.5-24B-A2B for knowledge-intensive work should pair it with retrieval or web search. Similarly, on coding benchmarks like SWE-Bench Verified and SciCode, larger models show stronger performance.

Inference speed makes on-device deployment practical. The model reaches 184 tokens per second on an Apple M5 Max and 114 tokens per second on a Ryzen AI Max+ 395, which is up to 1.7 times faster than competing sparse models while using less memory. On NVIDIA B300 GPUs at high concurrency, it achieves over 21,000 output tokens per second.

What to test

Before adopting LFM2.5-24B-A2B for production agents, verify several characteristics:

• Agentic reliability at scale. Liquid AI trained the model inside agent harnesses, but benchmark results reflect single-turn or limited multi-turn tasks. Test whether the model sustains reasoning quality over longer agent runs in your specific workflows.
• Tool-use accuracy. The model scores well on tool-use benchmarks, but tools vary widely. Build a test suite of your actual tools and verify the model calls them correctly and interprets their outputs as expected.
• Memory and latency in your environment. The reported speeds are on M5 Max, Ryzen AI Max+ 395, and B300 hardware. Measure actual performance on the CPUs and GPUs you plan to deploy on, especially under load or with batched requests.
• Context window utility. The 131,000 token context is a significant increase, but requires compatible inference backends. Test whether your chosen inference engine (llama.cpp, MLX, vLLM, SGLang) actually supports the full context in production and whether longer contexts degrade latency unacceptably.
• Reasoning tokens and cost. The model’s chain-of-thought reasoning adds tokens to every response. Understand your token budget under realistic agent workloads, since each thinking step consumes tokens even if the agent discards some reasoning.

The company notes day-one support exists across llama.cpp, MLX, vLLM, SGLang, and ONNX. Test with your chosen backend before production deployment.

Common questions

What are LFM models?

LFM stands for Liquid Foundation Models. They are a family of hybrid models designed for on-device deployment, combining gated short convolutions with grouped query attention. The LFM2.5 series extends this with reasoning capability and larger context windows.

Is LFM2.5-24B-A2B free to use?

Yes, both LFM2.5-24B-A2B and LFM2.5-24B-A2B-Base are open-weight models available on Hugging Face. There is no license fee to download, run, or fine-tune them, though you must follow the model’s license terms.

Where can I get LFM2.5-24B-A2B?

The model is available directly on Hugging Face. You can download the model weights, load it with transformers (version 5.0.0 or later), and run it locally with any compatible inference backend including llama.cpp, MLX, vLLM, SGLang, or ONNX.

The conclusion

LFM2.5-24B-A2B fills a practical gap: it is small enough to run on consumer laptops and desktops yet efficient enough at tool use and reasoning to power useful agents without cloud dependency. The sparse mixture-of-experts design keeps per-token costs low, making continuous agent operation feasible on limited hardware. Benchmark results show clear strength in agentic workflows and instruction-following, with documented weaknesses in knowledge-heavy tasks that are easily addressed through retrieval augmentation.

The open-weight release and day-one ecosystem support mean practitioners can start testing immediately. What matters next is real-world validation: whether the model maintains reasoning quality in multi-step agent loops specific to your domain, how tool-call accuracy compares to your actual tools, and whether the inference speed holds under production load.

Watch for community feedback on long-running agent deployments and for Liquid AI’s continued improvements to on-device reasoning. Multi-Agent AI Teams Burn Up to 5x Tokens for Little Gain, so efficiency gains from sparse models become increasingly meaningful as agent complexity grows.