Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model designed to run AI agents that call tools and complete multi-step tasks directly on a user's hardware. According to a blog post published on Hugging Face, the model has only 2B active parameters per token, yet it competes with a larger rival on agentic and tool-use benchmarks. The company says the model is available now for download, along with a base version.
A compact model built for agents
The headline claim is competitive performance from a model that is smaller than its peers. In the Hugging Face post, Liquid AI compares LFM2.5-24B-A2B with Qwen3.6-35B-A3B, which the company describes as having roughly 50 percent more total and active parameters. On the LiveClaw benchmark, the new model scored 48.09, compared with 38.19 for the Qwen model.
The model is also built to think before it answers. Liquid AI describes it as always reasoning, producing an explicit chain of thought before its final output. The company's argument is that, because only 2B parameters are active for each token, each reasoning step is comparatively cheap to run.
Liquid AI evaluated the model against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning across 16 benchmarks. These covered tool use and agentic workflows, knowledge, instruction following, math, and coding. The post states that the model is the smallest in that group but remains competitive with mixture-of-experts models that have about half again as many total and active parameters.
Where it leads and where it trails
According to the post, the strongest results are in tool use and instruction following. LFM2.5-24B-A2B leads or matches the competition on five of the eight tool-use and agentic benchmarks, including AutomationBench, BFCLv4, ToolSandbox, LiveClaw, and τ³-Bench Banking. On AutomationBench, the company reports a score more than three times that of the next-best model. It also leads on all three instruction-following benchmarks the company tested.
Larger models keep an edge on knowledge-heavy tests such as GPQA Diamond, HLE, and AA-Omniscience. They also lead on coding benchmarks including SWE-Bench Verified (OpenCode) and SciCode. Liquid AI recommends pairing the model with retrieval or web search for workloads that depend heavily on up-to-date or specialised knowledge.
How the model was built
LFM2.5-24B-A2B builds on the earlier LFM2-24B-A2B and keeps its core architecture. The model has 40 layers and 64 experts per mixture-of-experts block, with top-4 routing. It uses the same hybrid of gated short convolutions and grouped-query attention found in its predecessor.
Several changes distinguish the new release. The context window grows from 32K to 131K tokens. An in-place tokenizer expansion doubles the vocabulary from 65K to 128K tokens, which the company says improves tokenization for non-Latin scripts. Unlike its predecessor, LFM2.5-24B-A2B is a reasoning-only model.
Pre-training was scaled up from about 17 trillion to about 33 trillion tokens. A mid-training phase then added 2 trillion tokens at a 32K context length, weighted toward reasoning, math, code, instruction following, and multilingual text. A further 300 billion tokens at 131K context focused on long-context data. Tool calls and multi-turn trajectories appear throughout both phases.
Post-training into an agent
The post describes a five-stage post-training process that turns the base model into an agent. The first two stages use multi-domain on-policy distillation, known as MOPD. The first distills specialist teacher models on non-agentic tasks, and the second focuses on agentic tasks. The later stages rely on agentic reinforcement learning, in which the model is trained inside agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode.
Speed on laptops, workstations and servers
Liquid AI puts particular weight on inference speed. The company reports decoding 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395. According to the post, that makes it up to 1.7 times faster than the next-best compared model, while also using the least memory of the group.
On data-centre hardware, the company says the model reaches up to 21,000 output tokens per second at high concurrency on a single NVIDIA B300 GPU. Liquid AI describes it as the fastest model in its size class on that measure.
The model has day-one support across several inference tools, including llama.cpp, MLX, vLLM, SGLang, and ONNX. For developers who want to test it, the post says the latest version of the transformers library is required, specifically a release compatible with transformers 5.0.0 or later.
Running it in an agent harness
The model can be served locally behind any OpenAI-compatible backend, after which an agent harness is pointed at that endpoint. The post gives example port numbers that depend on the serving tool: 8080 for llama.cpp and MLX, 8000 for vLLM, 30000 for SGLang, and 1234 for LM Studio. Configuration details for OpenClaw and Pi are available on the model card, and Liquid AI points to a separate guide for the full setup.
- Download: both the base and instruction-tuned versions are on Hugging Face.
- Run locally: follow the guide for agents such as OpenClaw, Hermes Agent, and Pi.
- Fine-tune: adapt the model to a specific task using the company's tutorials.
What it means for developers and users
For developers building private assistants, the appeal is clear. A model that handles tool calls reliably and runs on a single consumer machine could reduce dependence on cloud APIs for some tasks. The trade-off the company itself acknowledges is weaker performance on knowledge-heavy and coding benchmarks, which means the model is better suited to agents that work with external tools and data sources than to tasks that rely on memorised facts.
The release also reflects a broader push to bring capable reasoning models onto local hardware. Whether the benchmark gains hold up in real deployments will depend on independent testing, which the company has not yet been reported to have received. Readers interested in the model can review the Hugging Face release for the full benchmark tables and setup instructions.
Related: Awesome Jev Catalogs 859 Projects for Fast Agent Decision Routing · AI Agents Decompiled a Shooter: Lessons From 500+ Billion Tokens
Comments (0)