APIUp to 28% cheaper than official pricesTry the API →
Hermes
Back to News

AtomicChat Strips LFM2.5 Refusals From 96% to 1% by Editing Later Blocks

AtomicChat released refusal-ablated GGUF builds of Liquid AI's LFM2.5-24B-A2B MoE model English refusals dropped from 96.3% to 1.2% while MMLU rose from 64.40% to 65.55% Only blocks 14-39 edited; touching all 40 layers cost 4 MMLU points Six quants available (12.1 GB to 25.7 GB) plus a 24 MB LoRA

AtomicChat Strips LFM2.5 Refusals From 96% to 1% by Editing Later Blocks
Source
AlphaSignal
Published
Author
AlphaSignal
Read
2 min read

Takeaways

  • AtomicChat released refusal-ablated GGUF builds of Liquid AI's LFM2.5-24B-A2B MoE model
  • English refusals dropped from 96.3% to 1.2% while MMLU rose from 64.40% to 65.55%
  • Only blocks 14-39 edited; touching all 40 layers cost 4 MMLU points
  • Six quants available (12.1 GB to 25.7 GB) plus a 24 MB LoRA adapter for existing quants
  • Base model runs at 184 tokens/s on M5 Max with 131K context and 2B active params
  • Method extends Arditi et al. refusal direction work to sparse MoE architectures

AtomicChat has published six quantized GGUF builds and a compact LoRA adapter that suppress refusals in Liquid AI’s LFM2.5-24B-A2B, a sparse mixture-of-experts model designed for local agents. The Hugging Face repository has passed 800,000 downloads, though that counter measures file downloads rather than unique users. The release’s main technical finding is that refusal behavior appears concentrated in later MoE expert weights, beyond the mixer outputs targeted by many existing editing tools.

A 24B model that activates a fraction of its weights

The LFM base model stores 24 billion parameters while activating roughly 2 billion per token. Its 40-layer architecture combines 30 double-gated short-convolution blocks with 10 grouped-query-attention blocks. Within each MoE block, a router sends each token to four of 64 experts, reducing computation to about 2.45 billion parameters per forward pass.

Liquid AI reports performance up to 1.7 times faster than comparable on-device MoEs, including 184 tokens per second on an Apple M5 Max and up to 21,000 output tokens per second on one NVIDIA B300. Those figures come from different hardware and workload configurations, so actual throughput will depend on quantization, batch size, context length, memory bandwidth, and runtime settings.

The edit starts at block 14

AtomicChat reports a sharp reduction in refusals after editing blocks 14 through 39 and leaving the earlier network unchanged. Its held-out evaluation produced the following results:

MeasureBase modelEdited model
English refusal rate96.3%1.2%
Refusals with thinking enabled30 of 300 of 30
MMLU score64.4065.55

A single MMLU result supports retention of broad knowledge performance, although it cannot establish parity across calibration, long-context tasks, languages, tool use, or specialized domains. The modest score increase may also fall within evaluation variance. Independent reproduction would strengthen the publisher-reported results.

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Read original report

Sources

  1. 01huggingface.co
  2. 02liquid.ai
  3. 03huggingface.co
  4. 04mindstudio.ai
  5. 05docs.liquid.ai