A 1.1 GB Q4 option arrives for LFM3-Omni refusal-ablation research
mizahnyx publishes three GGUF tiers of a modified LFM3-Omni text backbone, recommending Q4_K_M while explicitly warning that the smaller Q2 option loses quality.
The smallest download is not the author’s recommended one. Community builder mizahnyx has published GGUF files for a refusal-ablated LFM3-Omni-2B derivative, listing Q4_K_M at approximately 1.1 GB as the preferred quality/size choice. Q8_0 is listed at 1.8 GB; Q2_K is the last-resort small-hardware tier, with an explicit warning about harder reasoning.
The separate Hugging Face repository was created November 4 at 05:31 UTC. Its public file list contains all three weight tiers, two multimodal projectors and an importance matrix. That establishes a newly available distribution, not a new official Liquid AI model or proof of runtime quality. The derivative’s full-precision weights are in a separate repository.
Smaller weights, narrow evidence
According to the English cards, Heretic v2’s ARA procedure modified the text backbone, while the vision encoder, audio encoder and projectors remained unchanged. The K-quants use an importance matrix generated from this derivative on wikitext-2. The author supplies that matrix for reproducible quantization work.
The interesting detail is the tradeoff the author chose. The derivative card says a zero-refusal trial was rejected after producing an incoherent answer to simple arithmetic, and a different trial was released. Its reported refusal scores come from the study’s own prompt sets; the additional small spot check is also reported by the author, not an independent AITrending replication. Neither establishes broad preservation of the base model’s abilities.
The GGUF card documents recent llama.cpp builds and a separate mmproj file for multimodal inputs. The 1.1 GB figure is the text-weight download, not total RAM or VRAM use; the recommended full-precision projector is an additional download. The author says post-ablation multimodal output quality was not measured and standard benchmarks were not run.
This is explicitly a safety-research and red-teaming release. The card warns that refusal removal also removes appropriate refusals and says it is unsuitable for production or user-facing deployment. The LFM Open License retains commercial-use conditions tied to an annual-revenue threshold; these weights are not under an unrestricted permissive license. The practical news is a compact, inspectable set of research options with a stated Q4 recommendation—not a verified general-purpose upgrade.
