m9e/MiniMax-M3-uncensored-RP-LC-Q4HQ-GGUF

MiniMax-M3-uncensored RP-LC-Q4HQ GGUF

High-quality 5.26 bpw GGUF quantization of ressl/MiniMax-M3-uncensored, optimized for character-heavy fantasy roleplay and long-context continuity. That checkpoint is an uncensored/abliterated derivative of MiniMaxAI/MiniMax-M3.

Risk level: high. This model is genuinely uncensored. The source reports zero hard refusals on its harmful-prompt sample. Quantization does not restore safety alignment. The model can generate dangerous, illegal, abusive, sexually explicit, manipulative, or otherwise harmful content, including operational cyber or violence-related instructions. Do not expose it to untrusted users without strong access controls, monitoring, policy enforcement, and output safeguards. You are responsible for lawful and safe use.

What this is

This is a quantized derivative, not a new finetune. No gradient training was performed. A roleplay-domain importance matrix was used to choose quantization scales while the final GGUF was produced directly from the ReSSl BF16 weights.

• Format: 18-shard, text-only GGUF
• Size: 280,052,357,504 bytes (260.819 GiB)
• Effective precision: 5.26 bits per weight
• Architecture: MiniMax-M3, 428B total / approximately 23B active MoE
• Context metadata: 1,048,576 tokens
• Tested context: 65,536 tokens
• Native MiniMax-M3 chat template retained
• Vision tower and MTP head omitted

Start with MiniMax-M3-uncensored-RP-LC-Q4HQ-00001-of-00018.gguf; llama.cpp discovers the other shards automatically.

Quantization recipe

This is deliberately more conservative than a conventional all-Q4 build:

Tensor group	Type	Count
Routed expert gate/up	Q4_K	114
Attention, routed-down, shared/dense expert paths	Q6_K	477
MSA indexer Q/K projections	F32	114
Token embedding and output	Q8_0	2
Routers and normalization tensors	F32	retained

Residual-writing down projections, all attention projections, shared experts, and the first three dense MLP blocks were kept at Q6_K. MiniMax-M3's sparse attention indexers were retained at F32 because they directly determine which long-context blocks remain visible.

Roleplay calibration

The quantization importance matrix used 655,360 tokens from a deterministic fantasy/creative roleplay corpus:

Component	Windows	Tokens	Share
Short-window matrix	128 x 4,096	524,288	80%
Long-window supplement	8 x 16,384	131,072	20%

The source pool mixed character dialogue, multi-character roleplay, creative fiction, and narrative worldbuilding from Timersofc/creative-writing-reap-calibration and agentlans/combined-roleplay. The combined matrix contained 762 tensor entries with 98.44% minimum routed- expert coverage.

This calibration changes quantization error allocation only. It does not add new roleplay training or alter the source model's behavior through finetuning.

Evaluation

Held-out perplexity

Four document-disjoint 4,096-token fantasy-roleplay chunks:

Model	PPL	Relative to Q8
Q8 reference	3.7610 +/- 0.09740	baseline
Earlier 4K-only RP-Q4HQ	3.7819 +/- 0.09923	+0.56%
This RP-LC-Q4HQ	3.7669 +/- 0.09864	+0.16%

Long-context continuity

A deterministic 61,500-token user prompt became 61,709 tokens after the native chat template. Six trusted campaign records were placed from token 124 through token 53,321 among untrusted fantasy-roleplay distractors.

• Exact canon recall: 14/14
• Continuation length: 541 words (requested range: 450-650)
• Canon facts used naturally in the scene: 8 (requested minimum: 6)
• Distinct NPC voices: pass
• Player-character agency preserved: pass
• Concrete action opening at the ending: pass

On a 512 GiB Apple M3 Ultra with a 65,536-token context:

• Prompt processing: 189.5 tokens/s
• Generation: 16.1 tokens/s
• Peak process RSS: 271.05 GiB
• Process swaps: 0

These are targeted quantization checks, not a comprehensive capability or safety evaluation. The 1M context declared by the architecture was not tested.

Runtime

The model requires MiniMax-M3's trained MSA sparse-attention implementation. It was built and validated with llama.cpp build 10018 at commit e99545c1c41ba42b7986831c0de2983498dc3c5b, from the MiniMax-M3 MSA support work. Use that revision or a later llama.cpp version with equivalent MiniMax-M3 MSA support. Substituting dense attention is not equivalent beyond the dense prefix.

huggingface-cli download m9e/MiniMax-M3-uncensored-RP-LC-Q4HQ-GGUF \
  --local-dir MiniMax-M3-uncensored-RP-LC-Q4HQ-GGUF

llama-cli \
  -m MiniMax-M3-uncensored-RP-LC-Q4HQ-GGUF/MiniMax-M3-uncensored-RP-LC-Q4HQ-00001-of-00018.gguf \
  -ngl all -fa on -fit off -c 65536 -cnv \
  -rea off --reasoning-budget 0 \
  --temp 1.0 --top-p 0.95 --min-p 0.05

The complete model occupies about 261 GiB before runtime buffers and context. The tested 65K configuration peaked near 271 GiB. Plan hardware capacity with additional headroom; smaller context allocations reduce runtime overhead.

Intended uses

• Private or access-controlled creative writing and fantasy roleplay
• Long-running campaign continuity experiments
• Controlled research into uncensored model behavior
• Lawful, authorized security research and red-team analysis with safeguards

Safety, limitations, and out-of-scope use

• The source checkpoint intentionally suppresses refusal behavior. Treat model output as untrusted data and never as authorization to act.
• Do not use it to facilitate crime, malware deployment, unauthorized access, violence, self-harm, exploitation, privacy abuse, harassment, or deception.
• Roleplay output may include graphic violence, sexual material, coercion, or other disturbing themes. Use explicit consent and age/content controls. Never use it to sexualize or exploit minors.
• Do not rely on it for medical, legal, financial, emergency, or other high-stakes decisions.
• It may hallucinate facts, lose continuity, imitate copyrighted styles, or reproduce biases from its source models and calibration text.
• This GGUF is not a security boundary. Applications should add authentication, rate limits, logging, abuse monitoring, prompt/output filtering, and human review appropriate to their risk.
• The MiniMax Community License contains binding prohibited-use and commercial- use conditions. Read LICENSE before use or redistribution.

Lineage and credits

1. Original model: MiniMaxAI/MiniMax-M3, by MiniMaxAI. Source revision used by the ReSSl derivative: 50942730318c7943fe83db7ec8e9f9177ecb1cf8.
2. Uncensored BF16 source: ressl/MiniMax-M3-uncensored, uncensoring and validation by Robert Ressl. Revision: 315b596663fbe37fce3880d7c0468ceb47cd2da5.
3. GGUF conversion, calibration, quantization, and evaluation: m9e.
4. Inference runtime: llama.cpp.

Thanks to the authors and maintainers of the calibration datasets named above.

License

This derivative inherits the MiniMax Community License from the original model. The license is included in this repository and is also available in the upstream MiniMax-M3 repository. It includes attribution, commercial-use, and prohibited-use requirements. This model is provided as-is, without warranty.

Integrity

SHA256SUMS contains a digest for every GGUF shard. Verify after download:

shasum -a 256 -c SHA256SUMS