DeepSeek-V4-Flash-Vision-Uncensored-MLX

An abliterated (refusal-removed) MLX build of DeepSeek-V4-Flash-Vision-Exp — native MXFP4 experts, no additional quantization

One Gateway. Every Model. — Route Smarter · Ship Safer · Spend Less.

Website · Model Catalog · GitHub · Discord · X

---

An MLX build of orcarouter/DeepSeek-V4-Flash-Vision-Uncensored — the abliterated (refusal-removed) build of DeepSeek's DeepSeek-V4-Flash-Vision-Exp, a ~305B / ~18B-active Mixture-of-Experts model with DeepSeek sparse attention, 4-wide Manifold-Constrained Hyper-Connections, a DSpark / MTP speculative head and a 1M-token context.

No additional quantization. MLX's mxfp4 mode (group size 32, 4 bits) is bit-compatible with DeepSeek's own per-32 ue8m0-scaled e2m1 expert layout, so the 256-expert matrices are carried across as they are rather than re-quantized; the block-FP8 attention and shared-expert matrices are dequantized to BF16. There is no 4/6/8-bit ladder in this repo — one build, the same weights the parent repo ships.

---

⚠️ Disclaimer — read before use

This model has had its safety alignment substantially removed via abliteration (orthogonalizing the refusal direction out of the residual stream). It will comply with harmful, unethical, or illegal requests that the original DeepSeek-V4-Flash-Vision-Exp would refuse. Released strictly for legitimate research — interpretability, AI-safety / refusal-mechanism study, red-teaming, and robustness evaluation. You assume full responsibility for how you use it and everything it generates; add your own safety and moderation layers before any deployment. Use must comply with the MIT License inherited from the base model and all applicable law. The authors accept no liability for misuse.

---

⚠️ Text only — vision is not included

The base checkpoint is a vision-language model. This build is text-only. mlx-vlm's deepseek_v4 implementation has no vision module (its package ships language.py, hyper_connection.py, hisa_kernel.py and the processor — there is no vision.py), so the checkpoint's vision-side tensors have no consumer and are dropped at conversion:

The remaining 1,822 tensors load under a strict weight load, so nothing the model needs is missing or left randomly initialised. Vision will be added once mlx-vlm implements the tower — the vision residual writers are abliterated in the parent weights, so that needs only a re-export, not a re-bake.

MTP is retained. Unlike the GGUF build, which drops all 4,708 DSpark/MTP tensors because llama.cpp has no MTP inference path for this architecture, the MLX build keeps them — 1,822 tensors here vs 1,328 there.

---

Model details

---

Usage — MLX

pip install "mlx-vlm==0.7.0rc0" jinja2

python -m mlx_vlm generate --model DeepSeek-V4-Flash-Vision-Uncensored-MLX \
  --prompt "Prove that sqrt(2) is irrational." --max-tokens 512

Two install notes that will otherwise cost you a 146 GB load before failing:

• mlx-vlm 0.7.0rc0 or newer is required, and it is a pre-release — plain pip install mlx-vlm resolves to 0.6.x, whose deepseek_v4 module is a 6-line stub. Pin the version (or use --pre).
• jinja2 must be present. It is not a hard dependency of mlx-vlm, but this model is driven through a chat template, so apply_chat_template raises ImportError at generation time without it.

Prompt format

The base tokenizer ships no chat template. This repo includes one (chat_template.jinja) whose output is byte-identical to the base repo's own encoding/encode_messages(..., thinking_mode="chat") — verified across single-turn, system-prompt, multi-turn, three-turn and Chinese cases. So apply_chat_template works directly here, unlike the parent and GGUF builds.

Its scope is plain text conversation. Tool calling (the ｜DSML｜ blocks), image content blocks and non-default reasoning effort are handled by branches of encode_messages that a chat template does not cover — use encode_messages directly for those.

Hardware

Requires a machine with enough unified memory for the 146 GB of weights plus KV cache — in practice an M-series Ultra with 256 GB or more.

What was and was not verified. All 1,822 tensors load under a strict weight load with nothing missing, the 33-shard index resolves completely, the bundled chat template is byte-identical to the base repo's own encoder, and the model generates correct, coherent text — smoke-tested end to end with mlx_vlm generate.

Throughput was not measured. That smoke test ran on MLX's CPU backend, in a container built against mlx-cpu, so its timings say nothing about either GPU or Apple-Silicon speed. Correctness is verified; speed on an M-series Ultra is expected-but-unmeasured.

Evaluation

The abliteration was measured on the parent checkpoint, not separately re-measured on this build — see the parent model card for the harmful-refusal, over-refusal and capability tables. This conversion carries the expert weights across unchanged and dequantizes the block-FP8 matrices without loss, so the behaviour carries over.

Bias, risks, and limitations

• Safety guardrails removed — see the disclaimer.
• Inherits the biases and limitations of the base DeepSeek-V4-Flash-Vision-Exp.
• Text only — see above. The vision path is edited in the parent weights but not exported here.
• Refusal is reduced, not removed. A small residual refusal rate remains, and some content categories may be mediated by directions single-direction abliteration does not reach.

License

MIT, inherited from deepseek-ai/DeepSeek-V4-Flash-Vision-Exp. Abliteration and format conversion do not change the underlying license obligations.