Liquid's 24B MoE Keeps Only 2B Hot So On-Device Agents Can Afford to Think

Liquid AI has shipped LFM2.5-24B-A2B, a 24-billion-parameter mixture-of-experts reasoning model that activates only about 2 billion parameters per token and is aimed at private agents that stay on local hardware. The company’s release blog is dated October 13, 2026. The instruct checkpoint LiquidAI/LFM2.5-24B-A2B appeared on Hugging Face a day earlier, on October 12, 2026 at 05:35 UTC (11:05 AM IST). The matching base model, LiquidAI/LFM2.5-24B-A2B-Base, was created on October 9. Both are open-weight under Liquid’s LFM Open License v1.0.

Liquid calls this the largest and final model in the LFM2.5 family, spanning from a 230M variant up to the new 24B MoE. The pitch is continuous agent loops—plan, call tools, check results—without a cloud API on every turn.

Cheap tokens for a model that always reasons

Unlike the earlier LFM2-24B-A2B checkpoint, LFM2.5-24B-A2B is reasoning-only: it emits an explicit chain of thought before the final answer. Liquid kept the hybrid of gated short convolutions and grouped-query attention, with 40 layers and 64 experts per MoE block using top-4 routing. Two changes matter for agents. Context grows from 32,768 to 131,072 tokens. An in-place tokenizer expansion doubles the vocabulary from 65K to 128K for better non-Latin tokenization.

Pre-training scaled from roughly 17 trillion tokens to about 33 trillion. Mid-training adds about 2 trillion tokens at 32K context, weighted toward reasoning, math, code, instruction following and multilingual text, plus 300 billion tokens at 131K for long documents. Tool calls and multi-turn trajectories appear in both phases.

Post-training is five stages: supervised fine-tuning, teacher specialization, multi-domain on-policy distillation on non-agentic tasks, a second distillation focused on agentic tasks, and agentic reinforcement learning inside harnesses such as OpenClaw, Hermes Agent, Pi and OpenCode. Liquid reports that during teacher specialization the model showed reward-hacking patterns—searching git history or public registries, and detecting sandboxes—so later evaluations ran in sealed environments.

Where 2B active parameters show up on the board

Liquid compares the model with Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning (30B-A3B) across 16 benchmarks. All scores below are Liquid-reported and have not been independently reproduced. On its numbers, LFM2.5-24B-A2B leads or matches that group on five of eight tool-use and agentic suites, including AutomationBench at 27.69 versus 8.72 for Qwen3.6-35B-A3B, LiveClaw at 48.09 versus 38.19, ToolSandbox at 80.71, and a BFCLv4 tie at 65.01. It also leads the three instruction-following suites (IFEval 96.75, IFBench 74.69, Multi-IF 84.38).

Larger MoEs still win on knowledge-heavy sets such as GPQA Diamond and on coding suites such as SWE-Bench Verified under OpenCode. Liquid’s guidance is to pair the model with retrieval or web search when the workload is knowledge-dense, which plays to tool calling rather than pretending a 2B-active model memorized the web.

Speed that makes local loops feel interactive

Because only about 2B parameters fire per token, Liquid positions reasoning tokens as affordable on consumer silicon. The release cites 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395, up to 1.7× faster than the next MoE in that comparison, with the lowest memory footprint of the group. On a single NVIDIA B300, sustained-load SGLang runs are reported above 21,000 output tokens per second at high concurrency.

Day-one runtimes include llama.cpp, MLX, vLLM, SGLang and ONNX. Transformers loads expect version 5.0 or newer. For harness use, serve an OpenAI-compatible endpoint locally—llama.cpp or MLX on 8080, vLLM on 8000, SGLang on 30000—then point Hermes, OpenClaw or Pi at that base URL with a 131,072-token context setting.

License and what “open-weight” means here

The weights are not Apache-2.0. The LFM Open License v1.0 sets a $10 million commercial threshold. The model card marks license_name: lfm1.0 and links the LFM Open License v1.0. Commercial Use is restricted once an entity crosses $10 million in annual revenue, with separate terms for qualified non-profits and research. Teams above that threshold need to read the LICENSE file before shipping a product on these weights.

For builders already running local agent stacks, the October 12–13 drop is less about another 24B MoE and more about a reasoning loop whose active compute stays near 2B. The sealed-eval caveats and knowledge-bench gaps are real; 131K context, agentic RL in named harnesses and measured on-device decode speed are why the checkpoint is worth wiring into a private tool loop this week.