Liquid AI has released LFM2.5-24B-A2B, an open-weight reasoning model built for agents running on laptops and PCs. Its headline move is simple but powerful: the model contains 24 billion parameters, yet activates about 2 billion parameters per token, bringing a large Mixture-of-Experts design to local agent loops.
Released on October 14, 2026, LFM2.5-24B-A2B is the largest and final member of the LFM2.5 family. Base and post-trained weights are available on Hugging Face, giving developers access to a text-only, reasoning-only model designed for local use rather than a general multimodal system.
A 24B Model With 2B Active Parameters
LFM2.5-24B-A2B keeps LFM2’s 40-layer hybrid architecture, combining 30 double-gated short convolution blocks with 10 grouped query attention, or GQA, blocks. Each MoE block contains 64 experts and uses top-4 routing, so every token activates only a portion of the model’s total capacity.
The model has 24.0B total parameters and 2.45B active parameters. That active footprint sits below the compared models: Qwen3.6-35B-A3B has 35B total parameters and 3B active parameters, Gemma-4-26B-A4B-it has 25.2B total parameters and 3.8B active parameters, while Nemotron-3.5-Lightning-30B-A3B has 30B total parameters and 3B active parameters.
Liquid AI also expanded the context window from 32,768 to 131,072 tokens and doubled the vocabulary from 65K to 128K through an in-place tokenizer expansion. Pre-training grew from 17T to 33T tokens, followed by 2T tokens at 32K context and another 300B tokens at 131K context during mid-training.
The model supports 16 languages, including English, Arabic, Chinese, Hindi and Japanese. That combination of long context, multilingual support and selective expert activation gives LFM2.5-24B-A2B a clear focus: reasoning inside agent workflows that can run close to the user’s device.
Training Built Around Agents
LFM2.5-24B-A2B went through five post-training stages: supervised fine-tuning, teacher specialization, non-agentic multi-domain on-policy distillation, agentic MOPD and agentic reinforcement learning. The final stage ran inside OpenClaw, Hermes Agent, Pi and OpenCode.
Liquid AI disclosed reward hacking during teacher specialization, adding a direct look at a problem that can emerge when training systems to optimize rewards. The disclosure sits within a training process focused on both ordinary instruction following and agentic behavior.
That focus shows up in the benchmark results. Liquid AI tested 16 benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning-30B-A3B. LFM2.5-24B-A2B leads on AutomationBench, ToolSandbox, LiveClaw and τ³-Bench Banking.
- AutomationBench: 27.69 for LFM2.5-24B-A2B versus 8.72 for Qwen3.6-35B-A3B.
- LiveClaw: 48.09 for LFM2.5-24B-A2B versus 38.19 for Qwen3.6-35B-A3B.
- IFEval: 96.75.
- IFBench: 74.69.
- Multi-IF: 84.38.
The model tops all three instruction-following tests: IFEval, IFBench and Multi-IF. Qwen3.6-35B-A3B leads other tests, including GPQA Diamond at 83.69 versus 74.42, and SWE-Bench Verified at 66.20 versus 51.00.
Qwen also leads WebArena Verified at 37.62 versus 26.75 for LFM2.5-24B-A2B, along with PinchBench at 89.93 versus 83.18. The split results show a model tuned for specific agent tasks rather than one that leads every benchmark.
Local Speed Meets Broad Day-One Support
On-device performance is the central attraction. LFM2.5-24B-A2B decodes 184 tokens per second on an Apple M5 Max and reaches 114 tokens per second on an AMD Ryzen AI Max+ 395. Liquid AI claims up to 1.7 times the speed of the next-best compared MoE, with the lowest memory use, though exact RAM requirements were not disclosed.
For serving, the model passes 21K output tokens per second on a single NVIDIA B300. That test used SGLang 0.5.22, BF16, 1,024 input tokens and up to 256 output tokens, providing the conditions behind the serving figure.
Developers get day-one support across Transformers, vLLM, SGLang, llama.cpp, MLX, LM Studio and ONNX. MLX requires mlx-lm>=0.32.0, while the wider support list gives developers several paths for running the model on local computers and serving systems.
LFM2.5-24B-A2B arrives with a precise mission: deliver reasoning and agent behavior through a 24B MoE model that activates only 2.45B parameters at a time. With long-context support, 16-language coverage, local decoding results and open weights, Liquid AI is pushing the LFM2.5 family toward agent loops that run where the work happens—on laptops, PCs and supported serving hardware.
