LFM2.5-24B-A2B

// Overview

LFM2.5-24B-A2B is an open-weight reasoning model from Liquid AI, announced on the Liquid AI blog on October 13, 2026. It is a mixture of experts with 24.0B total parameters and 2.45B active per token (Liquid rounds this to about 2B, hence the A2B suffix), a 131,072-token context window and a 128,000-token vocabulary. Liquid describes it as the largest and final model in the LFM2.5 family, which runs from LFM2.5-230M up to this model, and it is a text-only model that always thinks before it answers: the chat template opens every assistant turn with a <think> tag.

It keeps the architecture of LFM2-24B-A2B (February 2026) — 40 layers, 30 of them double-gated short-convolution blocks and 10 grouped-query attention blocks, with 64 experts per MoE block and top-4 routing — and changes two things: the context window grows from 32K to 131K tokens, and the vocabulary doubles from 65K to 128K through an in-place tokenizer expansion for better non-Latin tokenization. Pre-training grew from about 17T to about 33T tokens, followed by 2T tokens of mid-training at 32K context and 300B tokens of long-context data at 131K. Post-training runs supervised fine-tuning, teacher specialization, two on-policy distillation stages (one non-agentic, one agentic) and agentic reinforcement learning inside the OpenClaw, Hermes Agent, Pi and OpenCode harnesses, so the model learns their tools and system prompts.

On Liquid's own 16-benchmark comparison against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it and Nemotron-3.5-Lightning (30B-A3B), it leads on AutomationBench (27.69), BFCLv4 (65.01), ToolSandbox (80.71), LiveClaw (48.09, against 38.19 for Qwen3.6-35B-A3B) and τ³-Bench Banking, and on all three instruction-following tests: IFEval 96.75, IFBench 74.69 and Multi-IF 84.38. Liquid is explicit that larger models lead on knowledge-heavy and coding work: Qwen3.6-35B-A3B scores 83.69 on GPQA Diamond to its 74.42, and 66.20 on SWE-Bench Verified (OpenCode) to its 51.00. The model card advises against using it alone for knowledge-heavy tasks and recommends pairing it with retrieval or web search.

Speed is the headline. On Liquid's device benchmarks (llama.cpp, Q4_K_M, 4K-token input) it decodes 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395 — up to 1.7× faster than the next-best compared model — while using the least memory of the four (15–16 GB). On a single NVIDIA B300 with SGLang and BF16 weights it passes 21K output tokens/s at high concurrency. Weights ship on Hugging Face in BF16, FP8, GGUF, ONNX and MLX packings alongside a separate base model, with support in Transformers, vLLM, SGLang, llama.cpp, MLX and LM Studio; Liquid Console serves it by API at $0.10 per 1M input tokens and $1.00 per 1M output tokens. Use is governed by the LFM Open License v1.0, whose commercial-use rights stop at a company annual revenue threshold of $10 million.