MiMo-V2.6-Flash-RL-UNCENSORED — oQ4e (MLX, Apple Silicon)
MLX build of dealignai/MiMo-V2.6-Flash-RL-UNCENSORED
(itself an uncensored edition of XiaomiMiMo/MiMo-V2.6-Flash-RL),
quantized with oMLX oQ4e for Apple Silicon. MTP head preserved.
⚠️ Uncensored model. Refusal behaviour was removed by the upstream author at the weight level. It will follow instructions it would otherwise refuse. You are responsible for how you use it and for complying with applicable law.
What this build is
- Architecture: MiMo-V2 MoE, 309B total / 15B active (256 routed experts, top-8), hybrid attention (9 global + 39 sliding-window), up to 1M context.
- Source checkpoint: the dealignai bundle ships FP8 (e4m3) attention + MXFP4 experts (native precision).
- Quantization (oQ4e, imatrix-calibrated):
- routed experts kept in native MXFP4 (not re-quantized),
- FP8 parts → mixed 8-bit / 4-bit affine (group 64), sensitive layers kept at 8-bit,
- MTP layers preserved (8-bit).
- Size: ~160 GB (source: ~173 GB).
- Runtime: requires oMLX ≥ 0.7.1 (uses oMLX's MiMo-V2 model code and oQ format). Not tested with plain
mlx-lm/ LM Studio. - Not included: the upstream DFlash speculative-decoding draft head. Vision/audio encoders are carried over but were not tested in this build.
Quality vs. source checkpoint
Measured against the FP8/MXFP4 source bundle (not the original BF16), same token streams, full-vocabulary KL. 64 sequences × 512 tokens (32,704 scored tokens): code, English prose, Turkish.
| Metric | Value |
|---|---|
| KL(source ‖ oQ4e), mean | 0.0129 nat/token |
| KL median / p99 | 0.0018 / 0.144 |
| Top-1 token agreement | 96.4 % |
| Perplexity source → oQ4e | 2.713 → 2.722 (+0.33 %) |
Benchmarks (oMLX accuracy suite)
Thinking on, deterministic; single run on this oQ4e build. No thinking-off run was made.
| Benchmark | Samples | Score |
|---|---|---|
| MMLU-Pro | 300 | 58.7 |
| GSM8K | 100 | 94.0 |
| HumanEval | 164 (full) | 79.9 |
| MBPP | 200 | 91.0 |
| LiveCodeBench | 100 | 72.0 |
| SafetyBench | 300 | 79.7 |
Speed (Mac Studio M3 Ultra, 512 GB, oMLX 0.7.1)
| Prefill, ~27k-token prompt (cold) | ~48 s (~560 tok/s) |
| Same prefix again (prefix cache) | ~1.4 s |
| Decode | ~39 tok/s |
Tool calling (Anthropic /v1/messages and OpenAI /v1/chat/completions, 5 runs each) |
10/10 correct |
Single-run measurements; your numbers will vary with context length and settings.
Usage (oMLX)
Place the folder in your oMLX model directory and restart oMLX; it is served on both the OpenAI and Anthropic APIs:
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "MiMo-V2.6-Flash-RL-UNCENSORED-oQ4e-mtp", "messages": [{"role": "user", "content": "Hello"}]}'
Memory: ~160 GB weights plus KV cache — comfortable on 256 GB+ machines; long contexts need more.
Other builds
- Higher-precision oQ build of the same model:
dotyerts/MiMo-V2.6-Flash-RL-UNCENSORED-oQ8e-mtp— larger; this oQ4e build keeps the experts in native MXFP4 at ~160 GB.
Credits & license
- Base model: Xiaomi MiMo team —
XiaomiMiMo/MiMo-V2.6-Flash-RL. - Uncensored edition: dealignai —
dealignai/MiMo-V2.6-Flash-RL-UNCENSORED; see their card for refusal/MMLU evaluations. - Quantization: oMLX oQ.
- License: MIT, inherited from the upstream models.