LFM2.5-24B-A2B

Overview

Liquid AI released LFM2.5-24B-A2B on October 13, 2026, an open-weight mixture-of-experts model with 24B total and 2B active parameters per token. It uses 40 layers with 64 experts per MoE block and top-4 routing, mixing gated short convolutions with grouped-query attention, a 131,072-token context window, and a 128K-token vocabulary, after pretraining on 33T tokens. Liquid AI reports it matches or beats Qwen3.6-35B-A3B on five of eight agentic and tool-use benchmarks (for example 48.09 vs 38.19 on LiveClaw) while larger models still lead on GPQA Diamond and SWE-Bench Verified. Speeds reported are 184 tokens/s on an Apple M5 Max and over 21K output tokens/s on a single NVIDIA B300 at high concurrency.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

• 24B total / 2B active parameters (MoE, 64 experts, top-4 routing)
• Context window extended from 32,768 to 131,072 tokens; vocabulary doubled to 128K
• Pretraining scaled from 17T to 33T tokens with a five-stage post-training pipeline including reinforcement learning
• Vendor-reported 184 tokens/s on Apple M5 Max and 114 tokens/s on AMD Ryzen AI Max+ 395
• Runs on llama.cpp (GGUF), MLX, vLLM, SGLang, and ONNX
• Base and post-trained weights on Hugging Face and the Liquid Playground

Use cases

• Tool-using agents running locally on laptops and workstations
• High-throughput agentic serving on a single data-center GPU
• Low-cost API workloads via Liquid Console

Pricing

Open weights (Liquid AI model license; commercial terms apply). API: $0.10 per 1M input tokens, $0.05 per 1M cached input, $1.00 per 1M output; $2 credit for new Liquid Console users.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released

13 Oct 2026

Record updated

18 Oct 2026

Record last checked

18 Oct 2026

• liquid.ai/blog/lfm2-5-24b-a2b
• orcarouter.ai/blog/lfm2-5-24b-a2b-release