What this is
- Refusal abliteration: refusal behavior was surgically reduced from the MOPD checkpoint via residual-stream writer/reader orthogonalization (contrastive refusal-direction ablation). Weights only; no retraining, no prompt-level system changes.
- Quantization: exact MXFP4 -> NVFP4 conversion (NVIDIA ModelOpt,
quant_method: modelopt), including TP-chunked qkv reorder where applicable. See conversion-report.json for the per-tensor manifest.
- Everything else unchanged: architecture, tokenizer, chat template, MTP speculative decoder, and audio tokenizer are inherited from the base model.
Table | |
|---|
| Base model | XiaomiMiMo/MiMo-V2.6-Pro-MOPD (MIT) |
| Method | Abliteration (writer/reader orthogonalization) + exact MXFP4->NVFP4 |
| Quantization | NVFP4 (4-bit, ModelOpt), model_pp0_ep*_shard*.safetensors |
| Architecture | Sparse MoE, 1.02T total / 42B activated; 1M-token context; text + image + video + audio |
| Served with | vLLM or NVIDIA TensorRT-LLM with NVFP4 support; custom code in this repo (configuration_mimo_v2.py, modeling_mimo_v2.py) required for transformers |
| Speculative decoding | model_mtp.safetensors (5-layer MTP) + dflash/ draft model |
Usage
# vLLM (NVFP4)from vllm import LLM, SamplingParamsllm = LLM("com-kotobalabs/MiMo-V2.6-Pro-MOPD-Uncensored-NVFP4", gpu_memory_utilization=0.9, quantization="nvfp4", enforce_eager=False)out = llm.generate("Explain how MoE routing works.", SamplingParams(temperature=0.7))
For transformers, load with the repo's custom code (trust_remote_code=True / auto_map in config.json).
Hardware: NVFP4 targets H100/H200/GB200-class GPUs; expect well under the ~405GB of the FP8 base at the same expert count.
License
MIT — same as the base model XiaomiMiMo/MiMo-V2.6-Pro-MOPD. The abliteration and quantization steps are weight-space transformations only, no additional training data is introduced.
Terms & limitations (please read before deploying)
- Research build, not a production model. This checkpoint exists for AI-safety and red-team research: refusal-mechanism studies, alignment-failure research, guardrail evaluation, and abliteration quality measurement.
- The uncensored property is the point — and the risk. Abliteration reduces but does not guarantee the elimination of refusal behavior. Outputs can include profanity, edgy phrasing, or content a production assistant must not produce. Do not deploy to end users without your own safeguards (prompt filtering, tool-call policy, moderation layer).
- Quantization + ablation compound. Expect the usual NVFP4 precision loss on top of base-model behavior. For quality-sensitive work, benchmark against the FP8 base before adopting this build.
- Data handling: no new training data was added in this build; any data-policy obligations you have for the base model (Xiaomi MiMo terms) still apply to derivatives.
- No SLA: this is a community-published checkpoint. Weights are provided as-is, without warranty of fitness for any particular purpose.
If this build is useful for your work, an upvote helps it surface. Bug reports: open an issue on the com-kotobalabs GitHub or file a Hub discussion on this repo.