X PAPER / Full editionIssue 010 · October 15, 2026

AI与科技 / 开放权重

A sparse 24B model targets local agents, with a revenue limit on commercial use

Source author: Gorden Sun · @Gorden_Sun
Source posted:
Written and edited by

Gorden Sun introduced Liquid AI’s LFM2.5-24B-A2B on October 14. Its model card specifies 24 billion total parameters and about 2.45 billion active per token. The “A2B” name is rounded; it does not mean memory only needs to hold two billion parameters. This text-only reasoning model supports tools and a 131,072-token context.

Native, FP8, GGUF, ONNX and four-bit Apple Silicon MLX versions offer different deployment options. The card recommends retrieval or search for knowledge-heavy work.

Liquid reports LiveClaw scores of 48.09 versus Qwen3.6-35B-A3B’s 38.19, but SWE-Bench Verified scores of 51.00 versus 66.20. That does not establish superiority across tasks. Its 184 tokens/second on M5 Max used llama.cpp, Q4_K_M quantization and a 4K input. Roughly 21,000 output tokens/second on B300 is aggregate high-concurrency throughput, not one response. We did not rerun the tests.

LFM Open License v1.0 conditions commercial rights on a $10 million annual-revenue threshold; commercial use above it is outside this license. The available weights are not covered by an unrestricted permissive license.

Sources and further reading

edited

Selected through a verified followed account: @Gorden_Sun.