Liquid AI releases LFM2.5 24B A2B for local agents
Liquid AI released LFM2.5 24B A2B on October 13, adding a larger reasoning model to its lineup for local agents. Both the base checkpoint and the version trained for agent tasks are available as downloadable weights.
What the model can handle
The model card lists 24 billion total parameters and 2.45 billion active parameters per token. Liquid rounds that active count to 2 billion in its launch announcement. The model accepts text and has a context window of 131,072 tokens. Its architecture combines short convolution blocks with attention and routes each token through four of the 64 experts in each expert block.
Download options include the native checkpoint, FP8, GGUF, ONNX and an MLX version for Apple Silicon. The DSpark draft model remains listed as coming soon. Liquid recommends pairing the model with retrieval or web search for tasks that depend heavily on factual knowledge.
Speed claims need context
Liquid reports 184 tokens per second on an Apple M5 Max and more than 21,000 output tokens per second on one NVIDIA B300. The GPU figure measures aggregate throughput under sustained concurrent requests. It does not describe one person receiving an answer at that speed. These are company benchmarks, and ByteForward has not independently tested the model.
Commercial use has a revenue condition
The weights use the LFM Open License v1.0. Its commercial use provisions include an annual revenue threshold of $10 million and exclude commercial use by entities exceeding that threshold. Businesses need to check those conditions before adopting it for paid products or internal commercial work. Qualified nonprofit organizations receive an exception to the threshold when using the model for research or other noncommercial purposes.
Redistribution also carries obligations. Recipients must receive the license, modified files need prominent notices stating that they were changed, and relevant copyright and attribution notices must remain. The release should be described as open weights with specific license terms.
Connecting the model to an agent
Liquid’s local agent guide describes a common setup for Hermes Agent, OpenClaw and Pi. Run a model server with tool calling enabled, then connect the agent to its endpoint using the OpenAI API format. Different serving tools use different default ports, so the agent configuration needs to match the server actually running.
Most examples in that guide still use Liquid’s smaller 2.6 billion parameter model. The guide says the same commands work with this release when the model identifier is replaced. Its smaller model memory figures should not be used to size a 24 billion parameter deployment. Quantization and the amount of context served also affect the memory tradeoff.
For a practical evaluation, start with the tasks the agent will actually perform. Check that it calls the intended tools, stays within the available memory and completes the work accurately before relying on its speed figures.
ByteForward’s reporting on Boole 1.7B examines how file format and runtime choices affect local deployment.
Illustrative laptop keyboard photographed by Yagnik Sangani. CC0 via the WordPress Photo Directory.
