Liquid AI LFM2.5-24B-A2B: 2B-active reasoning MoE with GGUF and MLX builds

Liquid AI released LFM2.5-24B-A2B on Oct 13, 2026. The blog gives 24B parameters with 2B active; the Hugging Face model card gives 24.0B total and 2.45B active. It is a mixture-of-experts model with 64 experts per MoE block and top-4 routing. It is reasoning-only: it writes an explicit chain of thought before the final answer. The model card calls it text-only. Compared with LFM2-24B-A2B, context grows from 32,768 to 131,072 tokens. A base model, LFM2.5-24B-A2B-Base, is released alongside. Liquid lists day-one support for llama.cpp (GGUF), MLX, vLLM, SGLang and ONNX. The model card also lists Transformers and LM Studio. Liquid's Hugging Face GGUF repo lists Q4_K_M at 14.53 GB, QAD-Q4_0 at 13.58 GB, Q5_K_M at 17.03 GB, Q6_K at 19.69 GB, Q8_0 at 25.49 GB, and BF16 and F16 at 47.96 GB each. Liquid says the QAD-Q4_0 file was trained to recover accuracy lost to Q4_0 quantization. Liquid also publishes MLX repos (4, 5, 6 and 8-bit, and bf16), plus ONNX and FP8 repos. The MLX 4-bit card says it needs mlx-lm 0.32.0 or later; earlier versions load the model but route the experts incorrectly. Liquid states decode speeds of 184 tokens/s on an Apple M5 Max and 114 tokens/s on an AMD Ryzen AI Max+ 395, run with llama.cpp at Q4_K_M and a 4K-token input. It also states over 21K output tokens/s on one NVIDIA B300 GPU with SGLang 0.5.22 in BF16. That is total throughput at high concurrency, not single-stream speed. These are vendor figures. We have not benchmarked the model. The weights use the LFM Open License v1.0, per Liquid's licence page and the Hugging Face LICENSE file. Commercial use by an organisation with $10 million or more in annual revenue is not covered; Liquid says it must buy a commercial licence.