Loading page…
LFM2.5-24B-A2B activates just 2 billion parameters per token. Liquid’s tests show strengths in tool use and instruction following, but larger rivals retain advantages elsewhere.
Listen to this story
LFM2.5-24B-A2B is designed for developers who want a tool-using reasoning model they can serve on their own hardware, rather than rely on a hosted API. Its mixture-of-experts design activates 2B of 24B parameters per token, but Liquid’s own comparisons point to a targeted strength—not a broad performance win: the model does well on tool use and instruction following while trailing a larger rival on web and coding tasks. For knowledge-heavy work, Liquid recommends pairing it with retrieval or web search.
The context window expands from 32K to 131K tokens, and the tokenizer vocabulary doubles from 65K to 128K.
Pre-training increased from roughly 17T to 33T tokens; later training included agent tasks and reinforcement learning in OpenClaw, Hermes Agent, Pi, and OpenCode.
Liquid reports that LFM2.5 led or matched rivals on five of eight tool-use and agent benchmarks and led all three instruction-following tests; these are company-reported results.
Liquid AI released LFM2.5-24B-A2B on October 13, 2026, giving developers downloadable weights for a reasoning model built to run agents on their own hardware. The model calls tools and handles multi-step tasks while activating only 2 billion of its 24 billion parameters per token. Both the trained model and its Base variant are available on Hugging Face.
The efficiency mechanism is a mixture-of-experts design, which routes work through selected parts of the model rather than activating everything. LFM2.5 keeps its predecessor’s 40-layer architecture, with 64 experts in each expert block and four selected at a time. The 2-billion active figure describes computation per token—not the model’s total size.
Unlike LFM2-24B-A2B, the new release is reasoning-only: it generates an explicit reasoning trace before its final answer. That behavior is part of the model’s design, rather than an optional mode described in the release. Liquid positions it for agents that must think through a task and call tools.
The context window—the amount of text the model can work with at once—grows from 32K to 131K tokens. Liquid also expanded the tokenizer’s vocabulary from 65K to 128K tokens, a change it says improves how the model represents text in non-Latin writing systems.
Training changed alongside the context limit. Liquid increased pre-training from roughly 17 trillion to 33 trillion tokens. Further training emphasized reasoning, math, code, instruction following and multilingual text, followed by long-context material. Tool calls and multi-turn task sequences appeared throughout those later training phases.
Post-training then moved through supervised examples, specialized teacher models and two stages of learning from those teachers: first on non-agent tasks, then on agent tasks. The final reinforcement-learning stage trained the model inside agent frameworks including OpenClaw, Hermes Agent, Pi and OpenCode.
The new release increases the context window while retaining its predecessor’s architecture.
Liquid evaluated the release against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning across 16 benchmarks. All models ran through the vLLM inference software using each benchmark’s recommended generation settings. These are company-reported results, not an independent evaluation.
The strongest results cluster around tool use and following instructions. Liquid’s model led or matched the comparison group on five of eight tool-use and agent benchmarks, and led on all three instruction-following tests. On LiveClaw, it scored 48.09 against Qwen’s 38.19.
Qwen retained a substantial advantage on other tasks. On WebArena Verified, Liquid scored 26.75 versus Qwen’s 37.62. On SWE-Bench Verified, evaluated through OpenCode, the scores were 51.00 and 66.20. The release therefore does not establish a general performance lead over the larger rival.
Liquid acknowledges that larger models retain an edge on knowledge-heavy and coding benchmarks. Its recommendation for knowledge-heavy workloads is to pair LFM2.5 with retrieval or web search. That positions the model as a tool-using component that can seek information, rather than a replacement for every larger model.
Day-one support includes llama.cpp, MLX, vLLM, SGLang and ONNX. Liquid reports decoding speeds of 184 tokens per second on an Apple M5 Max and 114 on a Ryzen AI Max+ 395. Those figures describe specific hardware tests, not a promised speed on every device.
The much larger GPU figure measures a different workload. Liquid reports more than 21,000 output tokens per second on one NVIDIA B300 at high concurrency—serving many requests at once. It should not be read as the speed of a single user’s local conversation.
For deployment, Liquid’s local-agent setup serves the model behind an OpenAI-compatible backend, then points the agent framework at that local address. Its Hermes example sets the model, context length and tool-use enforcement explicitly. Developers loading the model directly through Transformers need version 5.0.0 or later.
hermes config set model.provider custom
hermes config set model.base_url http://localhost:8080/v1
hermes config set model.default LFM2.5-24B-A2B
hermes config set model.context_length 131072
hermes config set model.api_mode chat_completions
hermes config set agent.tool_use_enforcement trueThe address must match the serving software: Liquid lists port 8080 for llama.cpp and MLX, 8000 for vLLM, 30000 for SGLang and 1234 for LM Studio. The model card also provides configurations for OpenClaw and Pi.
Loading discussion...
Join the conversation
Explain which matters more for the tasks you would delegate.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.