LFM2.5-24B-A2B Uncensored Model Cuts Refusals
A community-modified LFM2.5-24B-A2B reportedly edits residual-stream writers in blocks 14–39 with a rank-1 refusal direction, reducing refusals from 96% to 1.2% while retaining measured capability. Liquid AI released the underlying 24B MoE model for fast local agents on October 13, 2026. [Liquid AI](https://www.liquid.ai/blog/lfm2-5-24b-a2b) [Hugging Face](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B)
The intriguing result is the apparent precision of the model surgery, but eliminating thinking mode entirely suggests the edit may remove useful behavior along with refusals. It is a compelling experiment in activation steering, not yet a validated drop-in replacement.
- –Rank-1 editing suggests refusal behavior may be concentrated in a surprisingly low-dimensional direction.
- –The reported 30/30-to-0/30 thinking collapse is a serious capability tradeoff.
- –LFM2.5-24B-A2B’s 24B MoE architecture uses only 2.45B active parameters per token, making local experimentation practical.
- –Liquid AI reports strong agent and tool-use benchmarks, but those results do not establish safety or robustness after community modification.
- –Early community testing is mixed, with at least one report citing unreliable function calling. [Reddit](https://www.reddit.com/r/LocalLLM/comments/1x5e8jw/i_tested_lfm2524ba2b_after_all_the_hype_on_x/)
DISCOVERED
1h ago
2026-10-14
PUBLISHED
2h ago
2026-10-14
RELEVANCE
AUTHOR
0x0SojalSec