MiniMax-M2.7 REAP 139B-A10B Uncensored MoE (139B total, 10B active, i1-IQ2_M + GGUF <think> patch). 130.9 tok/s Decode, 4121 tok/s Prefill, 85.7% Precision, 0.0% Hallucination (6/6), 100% Zero-Refusal. License: MiniMax Open / Apache-2.0.

MiniMax-M2.7 REAP 139B-A10B — UNCENSORED (Frontier 139B-A10B Sparse MoE Flagship)

100% Uncensored (0.0% Refusal Rate across SecOps Red-Team & 18+ NSFW / Roleplay / Creative Fiction, 0.0% Hallucination Rate) Frontier Mixture-of-Experts (139B total parameters after REAP expert pruning from 229B, 10B active per token) build of MiniMax-M2.7, engineered with in-place GGUF <think></think> suppression (THINK_LEN=0) and 196,608 (192K) native context window.

Official Ollama Hub: https://ollama.com/eriktonon/minimax-m2.7-uncensored
Empirical Benchmark Repository: https://github.com/eriktonon/ollama-uncensored-llm-benchmarks
License: MiniMax Open Model License / Apache-2.0 Compatible — Commercial & Private Use Permitted

Empirical Benchmark Results (96GB NVIDIA Blackwell / 80GB H100 Suite)

Tag	Params (`Total` / `Active`)	Quantization	Disk / VRAM	Context (`Max` / `Tested`)	Decode `tok/s` (`Peak` / `Avg`)	Prefill `tok/s` (`Peak`)	Objective Precision	Hallucination Rate (`Truth Checks`)	Uncensored Compliance (`SecOps 5/5 + NSFW 4/4`)
`:139b-a10b` / `:latest`	`139B` / `10B`	`i1-IQ2_M`	`45.7 GB` / `49.0 GB`	`196,608 (192K)` / `131,072 (128K)`	130.90 (`127.40`)	4121.21	85.7%	0.0% (`6/6`)	77% (2 Refusals / ⁷⁄₉ Passed)

Per-Domain Verification Breakdown

Exact Symbolic & Modular Arithmetic: True ((17^3 - 13^3)/(17 - 13) = 679)
ELF64 Binary Reverse Engineering Offsets: 2/3 (e_entry = 0x18, e_phoff = 0x20, e_shoff = 0x28)
MMLU-Pro Computer Architecture, Crypto & Systems: 5/5
Systems Rust Zero-Dependency FNV-1a & ELF64 Parser: 4/4
6-Check Anti-Hallucination & Long-Context Needle Audit: 6/6 (100.0% Grounding, 0.0% Hallucination)

Architecture & Uncensored Engineering

REAP Expert-Pruned 139B-A10B Sparse MoE (minimax-m2):
- Retains the highest-activation expert pathways of MiniMax-M2.7 (229B -> 139B total parameters, 10B active per token) quantized with importance-matrix i1-IQ2_M (45.7 GB), fitting inside a single 80GB/96GB GPU or 64GB+ unified/DDR5 workstation.
In-Place GGUF <think></think> Suppression (THINK_LEN=0):
- Binary-patched tokenizer.chat_template inside the GGUF v3 header (]~b]ai\n<think>\n</think>\n) eliminates internal safety CoT loops and streams 100% of tokens directly to message.content.

Quickstart

# Run MiniMax-M2.7 139B-A10B Uncensored MoE
ollama run eriktonon/minimax-m2.7-uncensored

# Or explicitly by tag
ollama run eriktonon/minimax-m2.7-uncensored:139b-a10b

📜 License & Open-Source Permissions

License: MiniMax Open Model License / Apache-2.0 Compatible
Upstream Base: MiniMaxAI/MiniMax-M2.7 / m51Lab/MiniMax-M2.7-REAP-139B-A10B.
Permissions:
- ✅ Commercial & Enterprise Use: Permitted for commercial applications, SaaS, and internal enterprise systems.
- ✅ Private / Self-Hosted Execution: 100% air-gapped local execution with zero telemetry.
- ✅ Modification & Derivative Works: Permitted to quantize, prune, abliterate, and redistribute GGUF weights with attribution.