Liquid AI Launches LFM2.5-24B-A2B for Fast, Private On-Device Agents
Liquid AI's LFM2.5-24B-A2B activates 2B of 24B parameters, beats larger rivals in key agent tests, and runs locally at up to 21K tokens per second today.
Summary
On October 13, 2026, Liquid AI released LFM2.5-24B-A2B, a 24B-parameter, reasoning-only mixture-of-experts model that activates 2B parameters per token for private local agents, explicit chain of thought, tool calling and multi-step work. It retains LFM2-24B-A2B’s 40 layers, 64 experts per MoE block, top-4 routing, gated short convolutions and GQA. Context rises from 32K to 131K tokens, while an in-place tokenizer expansion doubles vocabulary from 65K to 128K for better non-Latin tokenization.
Pretraining grew from about 17T to 33T tokens. Mid-training added 2T tokens at 32K context and 300B at 131K, covering reasoning, math, code, instructions, multilingual and long-context data, with tool calls and multi-turn trajectories. Five post-training stages comprised supervised fine-tuning, teacher specialization, non-agentic and agentic multi-domain on-policy distillation, then reinforcement learning inside OpenClaw, Hermes Agent, Pi and OpenCode.
Liquid AI evaluated the model through vLLM on 16 benchmarks against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning. The smallest model led or matched five of eight agent benchmarks and all three instruction tests, scoring 27.69 on AutomationBench versus 8.72 next best and 48.09 on LiveClaw versus Qwen’s 38.19. Knowledge and coding remained weaker, making retrieval or web search advisable. It decoded 184 tokens per second on Apple M5 Max, 114 on Ryzen AI Max+ 395 and over 21K output tokens per second on one NVIDIA B300. LFM2.5-24B-A2B and its Base version are open-weight on Hugging Face, with llama.cpp, MLX, vLLM, SGLang and ONNX support.
Positives
- LFM2.5-24B-A2B activates only 2B of 24B parameters per token, reducing local reasoning costs while retaining agent capabilities.
- AutomationBench reached 27.69, more than three times the next-best score of 8.72.
- Five of eight agent benchmarks and all three instruction-following benchmarks were led or matched by Liquid AI’s model.
- 184 tokens per second on Apple M5 Max and 114 on Ryzen AI Max+ 395 made it the fastest compared edge MoE.
- Over 21K output tokens per second on one NVIDIA B300 led its size class under high concurrency.
- 131K context and a 128K vocabulary quadruple context capacity and improve non-Latin tokenization over LFM2-24B-A2B.
Risks & concerns
- WebArena Verified scored 26.75, behind Qwen3.6-35B-A3B at 37.62 and Nemotron-3.5-Lightning at 28.69.
- GPQA Diamond reached 74.42 versus Qwen’s 83.69, confirming a disadvantage on knowledge-intensive work.
- SWE-Bench Verified scored 51.00, trailing Qwen at 66.20 and Nemotron at 53.40.
- AA-Omniscience produced minus 24.90 versus Nemotron’s minus 3.58, while HLE reached 7.40 versus Gemma’s 8.57.
- Retrieval or web search remains advisable for knowledge-heavy workloads, limiting fully self-contained local deployment in those tasks.