Skip to main content
Back

dealignai Runs Uncensored DeepSeek V4 Flash on a MacBook at 90% HumanEval

An abliterated MLX build of DeepSeek V4 Flash squeezes a 284B MoE into 95GB, running unrestricted at 100k context on a 128GB Mac.

Model, Full Access
  • dealignai released an abliterated MLX quant of DeepSeek V4 Flash 0731 at ~95GB on disk
  • Runs on 128GB Apple Silicon Macs with ~100k practical context and MIT license
  • MMLU 71.84% vs 70.96% base, HumanEval pass@2 ties base at 90%
  • 100% compliance on 240 HarmBench real-harm behaviors, copyright refusals preserved intentionally
  • JANG affine format uses 8-bit attention/shared experts, 2-bit routed experts, ~15% slower than base
  • Positioned for security research and red-teaming via vMLX runtime or jang_tools

A group calling itself dealignai has published an uncensored, quantized port of DeepSeek's newest flagship coder that runs entirely on a single Apple Silicon machine. The release, posted on Hugging Face as DeepSeek-V4-Flash-0731-JANG-CRACK, is a mixed-precision MLX build of DeepSeek's July production model, stripped of its refusal training and squeezed into roughly 95 GB on disk.

DeepSeek V4 Flash 0731 is a sparse mixture-of-experts model with 13B active parameters out of 284B total, and the release focused entirely on post-training, signaling that improving intelligence is no longer just about scaling parameter counts. It outperforms DeepSeek-V4-Pro (Preview) on benchmarks despite its far smaller activated parameter count, and is broadly competitive with the strongest proprietary models available. What dealignai has done is take that model, remove the refusal direction from the residual stream, and repackage it so it fits on a MacBook.

A 95GB coder in your laptop

The headline claim from the release tweet is that you can now run a truly unrestricted and capable coder on a MacBook while still comfortably reaching 100k context at decent speeds. The mechanics are worth unpacking. The build uses a custom quantization format called JANG affine mixed-precision, where shared experts and attention are kept at 8-bit with group size 64, while routed experts drop to 2-bit at group size 32. That asymmetric layout preserves the parts of the model that matter most for accuracy while aggressively compressing the parts that are sparsely activated.

Hardware requirements are steep but not exotic:

  • Apple Silicon M-series Max or Ultra (M2 Ultra, M3/M4/M5 Max recommended)
  • At least 128 GB of unified memory, with peak load around 100 GB
  • At least 100 GB of free disk space
  • jang_tools 2.5.39 or later, since earlier versions have runtime bugs that misreport MMLU and slow generation roughly 6x