Compatibility
Supported models
Support means one repository at one exact commit, with recorded weights, tokenizer and chat template digests. A patch tested on that revision is not claimed to work on any other revision, size or quantization of the same family.
Registered base models 0
No base model is registered yet
Accepted adapter format
PEFT LoRA in safetensors, nothing executable.
Adapters ship as adapter_model.safetensors with an adapter_config.json. Pickled weights, merged checkpoints and packages with executable content are rejected. Patches that change the tokenizer or add special tokens are not accepted.
Adding a model
A checkpoint joins after a patch is tested on it.
A patch is trained and evaluated on the new checkpoint in the same harness, with base weights loaded from safetensors and remote code disabled. Each runtime is marked tested, untested or unsupported on the model page, and untested stays untested.
Closed API models
Claude, GPT and Gemini can't be patched.
A patch is a set of LoRA weights loaded next to the base model's own weights. Hosted API models don't expose their weights, so there is nothing to attach an adapter to. Retoque only supports open-weight checkpoints that you download and run yourself, and a Retoque registry key never works as a key for any model provider.