Liquid AI released LFM2.5-24B-A2B local Agent tests lead Qwen in some areas

Article

Liquid AI Upgrade 24B Local model: only activates 2B parameters,Agent tests beating larger Qwen

Beating AI news flash, cutting-edge on-device AI R&D company Liquid AI has released LFM2.5-24B-A2B, focusing on improving the local Agent model's reasoning and tool-calling capabilities. The previous generation LFM2-24B-A2B primarily emphasized runtime efficiency. The new version retains the 24000000000 total-parameter design, activating only approximately 2000000000 parameters per use, but its pretraining data has increased from 17000000000000 to 33000000000000 Token, while the context has expanded from 32K to 131K.

The new model reasons before answering and undergoes reinforcement learning directly in frameworks such as OpenClaw, Hermes Agent、Pi and OpenCode, practicing continuous tool calls, processing returned results, and completing multi-step tasks. In the official demonstration, a MacBook does not need to call a cloud model Agent, enabling Hermes API to search for trending projects on Hugging Face, compile a briefing, and continue operating based on follow-up questions.

Official tests show that on Apple's Agent Max, the new model can generate M5 per second, with a peak speed reaching 184 Token, or MoE times that of comparable 1.7 models. In the Agent test, LiveClaw scored 48.09, higher than Qwen3.6-35B-A3B’s 38.19; AutomationBench scored 27.69, more than 3 times Qwen’s score. However, in the SWE-bench Verified programming test, it scored only 51%, below Qwen’s 66.2%.

Original article link https://m.theblockbeats.info/flash/371760

Key points

Liquid AI has released LFM2.5-24B-A2B。
model retains 24000000000 total parameters, activating only approximately 2000000000 parameters each time.
Pretraining data increased from 17000000000000 Token to 33000000000000, while the context expanded from 32K to 131K。
The model undergoes reinforcement learning in frameworks such as OpenClaw, Hermes Agent、Pi and OpenCode,Agent among others.
Official tests show that on Apple's M5 Max, the model can generate 184 Token per second, with a peak speed reaching MoE times that of comparable 1.7 models.
LiveClaw test score: 48.09, higher than Qwen3.6-35B-A3B’s 38.19; AutomationBench score: 27.69, more than 3 times Qwen’s.
The model scored SWE-bench Verified programming test: 51%, below Qwen’s 66.2%。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a model designed for local devices: it has 24000000000 parameters in total, but activates only approximately 2000000000 parameters each time, so its focus is improving runtime efficiency. It was also specifically trained to continuously call tools, process results, and complete multi-step tasks.

Why it matters to readers

If the test results reported in the article hold, this model could improve the ability of local devices to run Agent and compete with larger Qwen models in some Agent tests.

It demonstrates that AI models do not depend on the cloud API and can still perform multi-step tasks such as searching, organizing information, and continuing operations on local devices.

Risks and unknowns

The performance and speed figures in the report come mainly from official tests and have not yet undergone independent verification.
The evidence does not provide a direct Liquid AI official announcement, model weights, or a complete evaluation methodology.
Different hardware, prompting conditions, and task types may affect the comparison between the model and Qwen.

Liquid AI released LFM2.5-24B-A2B local Agent tests lead Qwen in some areas

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center