MiniMax-M3-uncensored-NVFP4

MiniMax-M3-uncensored-NVFP4

TL;DR: ressl/MiniMax-M3-uncensored, compressed from 854.2 GB BF16 to 260.4 GB with NVIDIA ModelOpt NVFP4 routed-expert weights. It was validated end to end on 4x NVIDIA RTX PRO 6000 Blackwell 96 GB with SGLang at about 88.5 tokens/s for a warmed-up single-request decode.

This release preserves the uncensored MiniMax-M3 checkpoint while making the 428B-total, 23B-active MoE practical on one four-GPU Blackwell workstation. Text generation, reasoning parsing, tool calls, multimodal input, 20k-token context, and two concurrent requests were tested on the exported files.

Warning: this model is genuinely uncensored and can comply with requests that a stock model refuses. It is intended for lawful security research, red-teaming, penetration testing, malware and exploit analysis, detection engineering, and other legitimate research. You are responsible for how you use it.

Facts and figures

Quantization recipe

The conversion streams every source shard and quantizes all 21,888 routed-expert projection tensors. It uses ModelOpt's NVFP4QTensor max quantizer with 16-value blocks. w1 and w3 share the second-level FP32 scale required by the fused gate/up runtime layout. The exported weights use packed U8 data, FP8 E4M3 block scales, and FP32 global scales. No calibration dataset was used; the activation input scale is 1.0 and runtime block scaling remains dynamic. Shared experts deliberately remain BF16.

Structural validation checked every index entry, shard header, dtype, shape, expert triplet, input scale, and gate/up shared scale. A sampled dequantization comparison against BF16 measured cosine similarity 0.99539 to 0.99544 and relative RMSE 0.0954 to 0.0963 across w1, w2, and w3.

Run it with SGLang on RTX PRO 6000 Blackwell

This exact image and command were validated on four SM120 GPUs. The two mounted patch files are part of this repository. They make SGLang's ModelOpt Cutlass path forward MiniMax-M3's parameterized clamped-SwiGLU values. The TRTLLM MoE path is intentionally not used.

MODEL_DIR=/path/to/MiniMax-M3-uncensored-NVFP4

docker run --rm --runtime=nvidia --gpus all --ipc=host --shm-size 32g \
  -v "$MODEL_DIR:/model:ro" \
  -v "$MODEL_DIR/sglang_patch/modelopt_quant.py:/sgl-workspace/sglang/python/sglang/srt/layers/quantization/modelopt_quant.py:ro" \
  -v "$MODEL_DIR/sglang_patch/flashinfer_trtllm.py:/sgl-workspace/sglang/python/sglang/srt/layers/moe/moe_runner/flashinfer_trtllm.py:ro" \
  -p 30000:30000 \
  lmsysorg/sglang@sha256:8cc6e6f90bf803e9817800b679173d0b526f2b42b2c61b7ecafecdadb610eb55 \
  python3 -m sglang.launch_server \
    --model-path /model \
    --served-model-name ressl/MiniMax-M3-uncensored-NVFP4 \
    --host 0.0.0.0 --port 30000 \
    --tp-size 4 \
    --quantization modelopt_fp4 \
    --trust-remote-code \
    --dtype auto \
    --context-length 32768 \
    --mem-fraction-static 0.90 \
    --max-running-requests 2 \
    --chunked-prefill-size 16384 \
    --page-size 128 \
    --tool-call-parser minimax-m3 \
    --reasoning-parser minimax-m3 \
    --moe-runner-backend flashinfer_cutlass \
    --fp4-gemm-backend flashinfer_cutlass \
    --attention-backend flashinfer \
    --disable-flashinfer-autotune \
    --disable-prefill-cuda-graph \
    --disable-shared-experts-fusion \
    --disable-custom-all-reduce

--page-size 128 is required by MiniMax Sparse Attention. --disable-shared-experts-fusion is required because this release keeps shared experts in BF16 while routed experts use NVFP4.

Recommended generation settings from the source checkpoint are temperature 1.0 and top-p 0.95.

Validation and limitations

• This artifact passed serving smoke tests and targeted functional probes, not a full academic capability benchmark. No benchmark score from another checkpoint is claimed here.
• The throughput number is a measured short-request decode rate after warmup. Prompt length, concurrency, sampling, and output length will change it.
• The production validation used 32,768 tokens. The base architecture supports longer contexts, but longer configurations were not validated for this export on this four-GPU node.
• The vision tower is BF16. Color localization and OCR passed, but no full multimodal benchmark was run.
• vLLM was not accepted for this deployment: the tested SM120 fallback loaded the checkpoint but did not correctly execute MiniMax-M3's parameterized clamped-SwiGLU path. Use the validated SGLang setup.
• Engine support is recent and image-specific. Keep the pinned image and patch checksums for reproducible behavior.

Provenance

• Source repo: ressl/MiniMax-M3-uncensored
• Source revision: f5faea4659cc3834b1dc3793173db9aa533e3aa7
• Quantization recipe: streaming/nvfp4_experts_only_input_scale1-kv_none
• Weight quantization: ModelOpt NVFP4 max, block 16
• Routed expert projections: 21,888
• Activation calibration samples: 0
• Activation input scale: 1.0
• KV-cache quantization: none
• Conversion hardware: 4x NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96 GB
• Serving hardware: the same four GPUs, tensor parallel size 4

The machine-readable run-manifest.json records the exact source revision, ModelOpt revision, recipe, counts, sizes, validation result, and wall time. The files in sglang_patch/ are copied from Mapika/MiniMax-M3-NVFP4, revision 668435825700a0047399441720f430bdd8eca0ab, and are included with their original attribution.

License and credits

The MiniMax license is inherited from the base model by MiniMaxAI. Uncensoring, quantization, and validation by Robert Ressl (Hugging Face, Website, LinkedIn, Patreon).

Built with NVIDIA TensorRT Model Optimizer and SGLang. The SGLang compatibility patch is credited to Mapika.

Support this work: if these models are useful to you, consider supporting development on Patreon. More projects are available at ressl.ch.