Liquid LFM2.5-24B-A2B is an open-weight 24B reasoning model built for developers who want fast, private AI agents that run entirely on their own hardware. Released by Liquid AI on October 13, 2026, it is the largest member of the LFM2.5 family and uses a mixture-of-experts design that activates only 2B parameters per token, so it stays interactive on a laptop while still reasoning through multi-step tasks.
Core Features
- 24B total / 2B active MoE with a 131,072-token context window, double the 32K window of its LFM2-24B predecessor.
- Explicit chain-of-thought reasoning before every answer, tuned for tool calling and long agentic loops.
- Native runtimes across Apple, AMD, Qualcomm, and Nvidia hardware via llama.cpp, MLX, vLLM, SGLang, and ONNX exports.
- Reaches 184 tokens/sec on an Apple M5 Max and over 21,000 tokens/sec on a single Nvidia B300 at high concurrency.
- Trained inside real agent harnesses such as OpenClaw, Hermes Agent, Pi, and OpenCode to cut tool-dispatch errors.
Use Cases and Best For
- Privacy-sensitive enterprises running document review or signal routing without sending data to the cloud.
- Developers building always-on local assistants on MacBooks or workstations.
- Teams that need many parallel agent threads at zero marginal API cost.
Pricing
The weights are open-weight and free to download under the LFM Open License v1.0, with commercial use limited to entities under US$10M annual revenue. Hosted inference on Liquid Console costs $0.10 per million input tokens ($0.05 cached) and $1.00 per million output tokens, and new accounts get $2 in starter credit.
Our Take
Best for teams that need frontier-class reasoning without cloud lock-in; the trade-off is that you supply the GPU or accept Liquid's per-token API price for convenience.
Explore more open weights in our AI Engine and Model directory.
