Liquid AI releases LFM2.5-24B-A2B with open weights for local agents
-oOdUn61bDRSykk4H951SoDOHzqoOe1.png)
Liquid AI has made LFM2.5-24B-A2B available for local agents. It has 24 billion parameters, with 2 billion active per token, and a context window of 131 072 tokens. According to the company, it generates 184 tokens per second on Apple M5 Max hardware.
Liquid AI has released LFM2.5-24B-A2B, a model designed for agents running directly on a device. Publicly available weights for both the fine-tuned variant and the base variant, LFM2.5-24B-A2B-Base, are on Hugging Face; the models are also available in the Playground service from Liquid AI. The model has a total of 24 billion parameters and activates 2 billion when processing each token.
Compared with LFM2-24B-A2B, the context window has expanded from 32 768 to 131 072 tokens. The new model produces an explicit reasoning process before the final answer. It has supported llama.cpp, MLX, vLLM, SGLang and ONNX since release. According to measurements from Liquid AI, it achieves a generation speed of 184 tokens per second on Apple M5 Max hardware and 114 tokens per second on AMD Ryzen AI Max+ 395 hardware.
According to Liquid AI, the model achieves comparable or better results than Qwen3.6-35B-A3B in five of eight tests of tool use and agent tasks. The company reports a score of 48.09 versus 38.19 on LiveClaw and 27.69 versus 8.72 on AutomationBench. However, the results do not show an advantage across all tasks: for example, on SWE-Bench Verified using OpenCode, the new model scored 51.00, while Qwen3.6-35B-A3B scored 66.20. These results were published by the developer.
Why it matters
The available weights allow the model to be run and fine-tuned locally. The larger context window provides more room for documents and the history of the agent's work. However, the specific task matters when choosing a model: the results published by the company show stronger performance on tool use than on some knowledge and coding tasks.
Release card
LFM2.5-24B-A2B
Liquid AI
- Specifications
- The model has a total of 24 billion parameters, of which 2 billion are active when processing each token.
- Context
- The context window is 131 072 tokens.
- Inputs
- The model processes text and generates text responses as well as tool calls.
- Availability
- The weights are publicly available on Hugging Face. The model can also be tried in the Playground service from Liquid AI. It has supported llama.cpp, MLX, vLLM, SGLang and ONNX since release.
- LiveClaw 48.09 In the agent test, the developer reports a score of 48.09 versus 38.19 for Qwen3.6-35B-A3B.
- AutomationBench 27.69 In the agent task test, the developer reports a score of 27.69 versus 8.72 for Qwen3.6-35B-A3B.
- SWE-Bench Verified (OpenCode) 51.00 In the coding test using OpenCode, the developer reports a score of 51.00 versus 66.20 for Qwen3.6-35B-A3B.
- Rychlost generování na Apple M5 Max 184 tokenů za sekundu The developer reports measurements of response generation speed on Apple M5 Max hardware.
- Rychlost generování na AMD Ryzen AI Max+ 395 114 tokenů za sekundu The developer reports measurements of response generation speed on AMD Ryzen AI Max+ 395 hardware.
- Running agents directly on a device.
- Agent tasks involving tool calls.
- Tasks requiring adherence to text instructions.
- For knowledge-intensive tasks, the developer recommends supplementing the model with retrieval.
- It trails Qwen3.6-35B-A3B on the SWE-Bench Verified and SciCode coding tests.
Compared with the previous model, LFM2-24B-A2B, it has a larger context window and always produces an explicit reasoning process before the final answer. According to Liquid AI, it is competitive with Qwen3.6-35B-A3B in five of eight tests of agent tasks and tool use. According to the developer, the larger models in the comparison retain an advantage on knowledge-intensive and coding tests.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
Two audiences, two different impacts
What this means
For individuals
An individual can try a local agent to work with their own material and take advantage of the larger context window compared with the previous model.
For a business
Development teams have another model for embedding agents directly into devices. Support for several inference tools makes it easier to choose a deployment method, but differences in results across tasks require evaluation on the specific business process.
DevelopmentCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.