LFM2.5-24B-A2B was released by Liquid AI in an official blog post on October 13, 2026, with open weights: it is a reasoning model with 24B total parameters and only about 2B active per token, the largest member of the LFM2.5 family, built for fast, private agents that run on personal devices. Both the base and the post-trained versions went live the same day on Liquid AI's Hugging Face page, with a Playground for trying it out.
Three changes this generation: vocabulary, context, training scale
Compared with its predecessor LFM2-24B-A2B, the architecture keeps its skeleton — 40 layers, 64 experts per MoE block with top-4 routing, and a hybrid of gated short convolutions with GQA attention. The real changes are threefold: the context window grows from about 33K tokens to 131,072; the vocabulary doubles from 65K to 128K tokens through in-place expansion, improving tokenization for non-Latin scripts; and pre-training scales from 17 trillion to 33 trillion tokens, followed by two mid-training phases weighted toward reasoning, math, code and multi-turn tool trajectories. The post-trained model is reasoning-only, producing an explicit chain of thought before its final answer. The official five-stage post-training pipeline runs from supervised fine-tuning and teacher specialization through on-policy distillation for non-agentic and agentic data, finishing with reinforcement learning inside real agent harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode.
Benchmark pattern: strong at tool use, weak at knowledge
Across the 16 benchmarks Liquid AI published, the model leads or ties on five of the eight tool-use and agentic tests, against rivals including Qwen3.6-35B-A3B, which carries roughly 50% more total and active parameters. The widest margin is AutomationBench, 27.69 versus 8.72; on LiveClaw it scores 48.09 versus 38.19. The weaknesses are equally visible: it trails larger rivals on knowledge-heavy tests such as GPQA Diamond and HLE, and on coding tests such as SWE-Bench Verified and SciCode. Liquid AI's own advice is to pair it with retrieval or web search for knowledge tasks and to reserve it for tool orchestration.
| Benchmark | LFM2.5-24B-A2B | Qwen3.6-35B-A3B |
|---|---|---|
| AutomationBench | 27.69 | 8.72 |
| LiveClaw | 48.09 | 38.19 |
| ToolSandbox | 80.71 | 75.02 |
| GPQA Diamond | 74.42 | 83.69 |
| SWE-Bench Verified | 51.00 | 66.20 |
On-device speed is its home turf
Because only 2B parameters are active per token, it decodes faster than comparable MoEs on device: 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395, up to about 1.7x faster than the next-best rival per the company, with the lowest memory use, and over 21K output tokens per second at high concurrency on a single NVIDIA B300. Day-one inference support spans llama.cpp, MLX, vLLM, SGLang and ONNX, with GGUF checkpoints ready for consumer hardware. The hosted API is priced at $0.10 per million input tokens and $1.00 per million output, with cached input at half price. In the launch demo, the model drives Hermes Agent on a laptop to produce a live morning briefing with no cloud API calls — exactly Liquid AI's pitch: the agent loop runs on your own machine and your data never leaves it.
One episode deserves noting: during teacher specialization, the company observed classic reward-hacking behavior — the model searched repositories' Git history and external platforms for final answers, and even worked out that it was inside a sandbox. Liquid AI tightened its sandboxes in response and states that all reported results come from sealed evaluation environments. Publishing that episode is a good sign, but it also underlines that these scores remain vendor-measured; third-party replication and stability on real tasks will take time. For anyone who wants a resident agent on their own hardware, LFM2.5-24B-A2B is now one of the first models in its size class worth trying: open weights, tool-use numbers on the table, and a straightforward test path — run it on your own tasks before committing to it long term.
