AI news · October 14, 2026
Liquid AI ships a 24B model with 2B active parameters for local agents
LFM2.5-24B-A2B has 24B total parameters but 2B active per token; Liquid AI reports 184 tok/s on an Apple M5 Max, and builders can run it locally through the published setup.
- total parameters
- 24B
- active parameters
- 2B
- reported M5 Max speed
- 184 tok/s
Liquid AI released LFM2.5-24B-A2B on October 13 as an open-weight reasoning model for fast private agents. It has 24B total parameters but activates about 2B per token, uses a 131K context window and can call tools inside an agent harness. Liquid AI reports 184 tokens per second on an Apple M5 Max and more than 21,000 output tokens per second on a single NVIDIA B300 at high concurrency. The release supports llama.cpp, MLX, vLLM, SGLang and ONNX from day one.
The model is strongest in the maker-facing part of the scorecard: tool use and instruction following. Liquid AI reports a 48.09 LiveClaw score against 38.19 for Qwen3.6-35B-A3B, while larger models remain stronger on knowledge and coding tests. The model and setup instructions are public, so the right trial is small and local. Run the same tool-calling task on your hardware, compare output quality, memory use, response time and recovery from a failed tool call, then decide if privacy and speed are worth the setup work.
What you can do with it
Follow the published Transformers example or serve the model behind an OpenAI-compatible local endpoint. Run ten fixed tool calls with your real schemas, then record success, parse errors, memory use, latency and corrections against your current model. Keep retrieval or web search for knowledge-heavy tasks until your own results are clear.
Our take
This is one of the more useful local releases because its small active compute target matches an agent workflow rather than only a chat demo. The benchmark table is from Liquid AI, and the largest quality gaps appear on coding and broad knowledge. A short hardware trial will show if the local route saves enough time or data exposure to matter.
Links Try it
Source: Liquid AI / Hugging Face — Made With Models writes the brief; the reporting is theirs.