Abliterated model forks like this are useful probes: they show how refusal and safety vectors can be extracted, adapted, and re-evaluated. Gemma 4 31B JANG_4M CRACK v2 is a practical example — a multimodal Gemma 4 variant retuned for conversational image-text tasks and stabilized thinking-mode behavior while reporting its safety and capability tradeoffs.
Key Capabilities
- Source & architecture: built from google/gemma-4-31b-it — dense 31B model with hybrid sliding/global attention, JANG v2 (MLX-native) safetensors format. (So what: compatible with MLX tooling and tuned for Gemma 4 internals.)
- Multimodal vision preserved: vision encoder kept in float16 for image-text inputs. (So what: retains visual understanding without extra conversion steps.)
- Abliteration & thinking-mode: uses CRACK v2 abliteration and provides recommended inference settings for thinking ON/OFF. (So what: enables chain-of-thought style reasoning with stability safeguards.)
- Benchmarks & compliance: reports 93.7% on a 300-prompt HarmBench and ~71.5% MMLU (200 subset). (So what: shows the model is permissive in some scenarios but still reports measured rates of refusal and capability.)
- Practical notes: model package is ~21 GB, tuned for vMLX and Apple Silicon workflows; license listed as 'gemma' on the card. (So what: modest hardware footprint for a 31B multimodal fork.)
Who it's for and tradeoffs
Great fit if you want a research/probing build of Gemma 4 that: explores safety/abliteration effects, experiments with thinking-mode chain-of-thought in multimodal settings, or integrates with vMLX for interactive vision+text experiments. Look elsewhere if you need a strictly production-reviewed, conservatively licensed model — abliterated forks can raise ethical, legal, or compliance concerns, and some use-cases (e.g., regulated clinical or safety-critical deployments) require models with explicit commercial/open licenses and full provenance. Users are responsible for legal and ethical compliance when running permissive/abliterated builds.
Where it fits
This is not the official Google Gemma release but a third-party, research-oriented fork hosted on Hugging Face by dealignai. Treat it as a research/demonstration artifact that surfaces tradeoffs between capability and refusal behavior rather than a drop-in enterprise model.
