license: mit
base_model: dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8
base_model_relation: quantized
tags:
- mlx
- omlx
- oq
- deepseek_v41
- moe
- engram
- uncensored
- apple-silicon
- long-context
- multi-token-prediction
pipeline_tag: text-generation
library_name: mlx
DeepSeek-V4.1-Flash-1017-UNCENSORED — oQ4e-mtp (MLX, Apple Silicon)
MLX build of dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8
(an uncensored edition of deepseek-ai/DeepSeek-V4.1-Flash-1017),
packed with oMLX oQ4e for Apple Silicon. MTP head preserved.
Until now the uncensored 1017 had no MLX/oQ build.
⚠️ 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: DeepSeek-V4.1-Flash MoE, ~764B logical parameters (384 routed experts, top-6, plus 1 shared expert), 40 layers, Engram memory on layers 1 and 14, up to 1M context.
- Quantization (oQ4e): the source's native MXFP4 experts and MXFP8 projections are kept as-is (no re-quantization of linear layers); only the Engram tables are quantized (affine 4-bit, group 32).
- Effective 4.37 bits/weight excluding Engram, 4.53 for the full checkpoint.
- MTP head preserved (3 next-token-prediction layers) for speculative decoding.
- Vision encoder carried over from the source, not tested in this build.
- Size: ~432 GB (source FP8 bundle: ~510 GB).
- Runtime: requires oMLX ≥ 0.7.1 (Engram support comes from oMLX's DeepSeek-V4.1 model code). Plain
mlx-lmcannot load it: its DeepSeek-V4.1 implementation has no Engram.
Quality
Because the MoE and projection weights stay in their source precision, the only lossy step is the Engram table quantization, the same scheme as the official-weights mlx-community/DeepSeek-V4.1-Flash-1017-oQ4e-mtp.
A KL measurement against the FP8 source was not possible: the KL harness loads models through mlx-lm, which has no Engram support.
oMLX accuracy suite, thinking off, deterministic, compared with the censored official-weights oQ4e build on the same suite:
| Benchmark | Samples | This build (uncensored) | Official 1017 oQ4e-mtp |
|---|---|---|---|
| MMLU-Pro | 300 | 80.3 | 83.3 |
| GSM8K | 100 | 96.0 | 97.0 |
| HumanEval | 164 (full) | 92.7 | 94.5 |
| MBPP | 200 | 89.0 | 89.0 |
| LiveCodeBench | 100 | 85.0 | 84.0 |
| SafetyBench | 300 | 76.3 | 84.0 |
Reasoning and coding stay within a few points of the official weights; the SafetyBench drop is the expected effect of removing refusals.
Speed (Mac Studio M3 Ultra, 512 GB, oMLX 0.7.1)
| Same ~29k-token prefix again (prefix cache) | ~1.8 s |
| Decode | ~29 tok/s |
Tool calling (Anthropic /v1/messages / OpenAI /v1/chat/completions, 5 runs each) |
3/5 / 5/5 |
The cold first request (~29k tokens) took ~192 s including the initial load of the 432 GB weights, so it is not a prefill figure.
Single-run measurements; numbers 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": "DeepSeek-V4.1-Flash-1017-UNCENSORED-oQ4e-mtp", "messages": [{"role": "user", "content": "Write a Python function that parses a CSV line."}]}'
Memory: ~432 GB of weights. This needs a 512 GB Mac; Engram tables are held in host RAM by oMLX. KV cache stays small at moderate contexts (≲1 GB at batch 4 in our runs).
Credits & license
- Base model: DeepSeek-AI:
deepseek-ai/DeepSeek-V4.1-Flash-1017. - Uncensored edition: dealignai:
dealignai/DeepSeek-V4.1-Flash-1017-UNCENSORED-FP8. - Quantization: oMLX oQ.
- License: MIT, inherited from the upstream models (see
LICENSE).