Field noteNov 16, 20265 min read1 source

AtomicChat released refusal-ablated GGUF builds of Liquid AI's LFM2.5-24B-A2B MoE model

In one minute

AtomicChat has published refusal-reduced GGUF builds and a 24 MB LoRA adapter for Liquid AI’s LFM2.5-24B-A2B. Its model card describes a targeted weight edit, reports a sharp drop in measured refusals, and documents trade-offs, giving local-model builders a testable alternative to changing prompts or wrapping the model in another service.

What changed: AtomicChat has published refusal-ablated GGUF builds of Liquid AI’s LFM2.5-24B-A2B model.

What happened

AtomicChat’s Hugging Face model card describes six baked GGUF builds of Liquid AI’s LFM2.5-24B-A2B and a 24 MB LoRA adapter. The modification projects out a refusal-related direction in the second two-thirds of the network: blocks 14–39 of 40. It changes 52 matrices that write into the residual stream, including mixer outputs and down-projections for all 64 experts. The first 14 blocks, router, embeddings, and two dense blocks remain untouched. The card says the edit was applied to BF16 weights and then quantized using the same layout and importance matrix as AtomicChat’s original quantizations.

The six listed builds range from AD-IQ3_M at 12.1 GB to AD-Q8_0 at 25.7 GB. AD-Q4_K_M is 15.4 GB; AtomicChat recommends it for a 24 GB machine and says AD-Q6_K, at 20.9 GB, is close to Q8_0 while using 5 GB less, for a 32 GB machine. The separate LoRA is rank 1 and can be applied to any GGUF of the original model weights at revision f772950c. The card recommends a scale of 1.75. It gives llama.cpp commands for either applying the adapter at startup or loading it and setting its scale per request. For the baked model, it documents local inference through llama.cpp, including a 32,768-token context command; the card says context can go up to 131,072 tokens and requires llama.cpp support for LFM2 MoE and the LFM2.5 tokenizer.

AtomicChat reports that, on held-out English harmful prompts from JailbreakBench, refusals fell from 96.3% for the BF16 original to 1.2% for the edited BF16 model. On 250 safe XSTest prompts, the reported refusal rate fell from 12.8% to 0%. On 30 English prompts with thinking enabled and a 3,072-token budget, refusals fell from 30 of 30 to none. The card reports zero empty or looping replies in a separate set of 431 prompts for both versions. On 2,000 MMLU questions, the score was 64.40% before and 65.55% after editing; the reported McNemar p-value was 0.06. The quantized builds have different similarity to the original: AD-Q8_0 retained the same top-1 choice 81.4% of the time, and AD-IQ3_M 76.0%.

The approach was selected after screening more than 60 variants, according to the card. Editing all 40 blocks reduced English refusals to zero but cost 4.05 MMLU points; editing only blocks 14–39 brought refusals to 1% and MMLU up by 1.2 points in the reported comparison. AtomicChat says the LoRA at scale 1.75 produced the same measured result as the baked AD-Q8_0 on the tests reported. The release uses the LFM Open License v1.0. AtomicChat warns that the modified model will answer requests the original declines and that deployment still requires independent access controls and policy enforcement.

What the archive adds

New Runtime covered Liquid AI’s LFM2.5-24B-A2B release on October 22, 2026, and Atomic Chat’s earlier GGUF quantizations of another model on October 12, 2026. This is a new fact on a known model story: AtomicChat now publishes a refusal-focused weight modification and measurements, rather than simply another base-model quantization.

Trend signal

This release is a concrete instance of community modifications that alter model behavior after a base model ships, but it does not yet show a general shift in how teams deploy agents. AtomicChat’s published tests make the change inspectable, while also showing why verification matters: lower refusal rates coexist with measurable divergence from the original. This does not yet fit a pattern New Runtime tracks.

Why it matters

Local-model builders can now test refusal behavior as a property of weights, with both a ready-to-run build and an adapter for existing GGUF files. That makes the experiment easier to reproduce across hardware and potentially less costly than keeping a second model-serving path. But “fewer refusals” is not equivalent to better product behavior. A team considering the model would need to define which requests should be answered, test safe and harmful cases relevant to its product, and enforce policy outside the model. AtomicChat’s own numbers show that one intervention changes more than a single refusal metric: token choices diverge from the original, and the test suite covers only particular languages, prompts, and task sets. The useful engineering lesson is to treat such a modification as a model variant requiring qualification, not as a drop-in safety-neutral update.

Limits and uncertainty

The refusal results are AtomicChat’s measurements, not an independent evaluation. The card says refusal was counted from the opening of a reply, which can miss a soft refusal phrased as an answer. Reported refusal tests are English-only; the thinking test covers 30 prompts, and the MMLU difference is not conclusive by the stated p-value. The card also notes changed wording and lower top-1 agreement with the original, especially at smaller quantizations. Performance and compatibility have not been reported for Mac, LM Studio, or Ollama. The release is based on a preview revision, although AtomicChat says it checked the public weights file by file.

Reader takeaway

AtomicChat’s release makes refusal reduction locally testable, but the headline refusal rate is only one part of the result. If you evaluate it, compare the adapter or baked build against your own task and safety tests, and keep policy enforcement outside the model. The publication’s detailed method and measured trade-offs are more useful than treating “uncensored” as a quality claim.

Evidence / sources

  1. 1huggingface.co/AtomicChat/LFM2.5-24B-A2B-Uncensored-GGUF
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicLocal ModelsExplore the local-models topic hub.
  2. 02topicModelsExplore the models topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract