Today for AI
HOT RADAR
llmHEAT 32.8°

AI CLUSTERED EVENT · 10/14/2026

Liquid AI Open-Sources LFM2.5-24B-A2B: MoE with 2B Active Params Achieves 184 Tokens/s on Edge

3 reports archived3 independent sourcesupdated 10/15/2026, 07:59:37
Synthesis & Latest Updates
3 Sources Cross-Validated

Liquid AI has open-sourced its flagship Mixture-of-Experts model, LFM2.5-24B-A2B, which achieves high-speed private agent deployment on edge devices by activating only 2B parameters per token out of a total 24B. This breakthrough significantly reduces inference latency and memory footprint, setting a new benchmark for privacy-sensitive local applications.

Cross-verified reports indicate that the model supports a 131K context window and outperforms larger models like Qwen3.6-35B-A3B in agentic benchmarks such as LiveClaw and AutomationBench. Thanks to its efficient sparse activation architecture, it reaches 184 tokens/s on an Apple M5 Max and up to 21K tokens/s on a single NVIDIA B300 GPU. Both base and post-trained weights are available on Hugging Face and integrated into OpenRouter.

This release marks a critical shift from 'usable' to 'highly efficient' edge AI. By addressing cloud API latency and data privacy concerns, Liquid AI empowers developers to build low-barrier, high-performance local autonomous agents, accelerating the adoption of complex reasoning tasks on edge computing devices.

LATEST/Base and post-trained weights for LFM2.5-24B-A2B are now fully available on Hugging Face and OpenRouter, enabling immediate local fine-tuning and integration into autonomous agent frameworks.

HEAT TRENDHourly heat curve

36 fully observed hours
Now
32.8
Peak
46.202:00
24h change
-9%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 3 reports · Latest first
  1. Hacker News AIT2·84 pts
    • LFM2.5-24B-A2B is the flagship of the LFM2.5 family, spanning from 230M to 24B parameters.
    • It scores significantly higher than Qwen3.6-35B-A3B on agentic benchmarks like LiveClaw and AutomationBench.
    • With only 2B active parameters, it achieves 21K tokens/s on a single NVIDIA B300 and 184 tokens/s on Apple M5 Max.
    • Both base and post-trained weights are open-source on Hugging Face, supporting local fine-tuning and deployment.
  2. OpenRouter (@OpenRouter)T1·78 pts
    • LFM2.5-24B-A2B uses an MoE architecture with 24B total parameters but only 2B active per inference, maximizing local efficiency.
    • The model supports a 131K token context window and a 128K vocabulary, specifically optimized for multi-step reasoning and tool use.
    • It achieves competitive performance against Qwen3.6-35B-A3B (which has ~50% more parameters) on agentic and tool-use benchmarks.
  3. Hugging Face BlogT1·78 pts
    • LFM2.5-24B-A2B utilizes a Mixture-of-Experts architecture deeply optimized for resource-constrained edge computing scenarios.
    • The model emphasizes speed and privacy, addressing pain points of high cloud API latency and data leakage risks.
    • Open-sourced on Hugging Face, allowing developers to directly integrate it for building locally running autonomous agents.