Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-MLX

DeepSeek-V4-Flash-Vision-UNCENSORED (Official Apple Silicon MLX)

Official Solstice-AI Apple Silicon Release • Native MLX MoE Engine • Lossless Multimodal Vision • Bundled Native MLX DSpark Speculative Drafter

Original Architecture by DeepSeek AI • Uncensored Weights by orcarouter • Apple Silicon Optimization by Solstice-AI

---

Model Summary

Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-MLX is the official Apple Silicon MLX release of DeepSeek-V4-Flash-Vision-UNCENSORED (284B total parameters, 256 routed MoE experts, ~13B active per token).

Features full native Apple Silicon Metal acceleration, preserving all 32-layer Vision Transformer (ViT) weights in pure BF16 and bundling the official native MLX DSpark speculative drafter in speculative/.

---

Official DeepSeek-V4-Flash Benchmark Scoreboard

Benchmark Suite	Discipline	DeepSeek-V4-Flash MLX	Claude 3.5 Sonnet	GPT-4o
Terminal-Bench 2.1	Agentic Terminal / CLI Execution	83.9%	63.5%	58.7%
SWE-bench Verified	Real-World Software Engineering	65.8%	61.2%	48.9%
LiveCodeBench v6	Competitive Algorithmic Coding	84.2%	78.4%	72.8%
MATH-500	High-School / Olympiad Math	94.6%	89.2%	91.4%
AIME 2025	American Invitational Mathematics Exam	78.2%	72.5%	63.8%
MMMU (Multimodal)	Multi-Discipline Visual Understanding	71.4%	70.4%	69.1%
DocVQA / ChartQA	Complex Document & Graph Reasoning	92.3%	91.8%	89.5%

---

Speculative Decoding Acceleration

This release bundles the official native Apple Silicon MLX DSpark drafter directly in speculative/:

• speculative/model.safetensors
• speculative/config.json
• speculative/tokenizer.json

Running with MLX-LM on macOS:

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load(
    "Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-MLX",
    draft_model="Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-MLX/speculative"
)
response = generate(
    model,
    tokenizer,
    prompt="Explain how Multi-Head Latent Attention works in DeepSeek-V4.",
    max_tokens=1024,
    verbose=True
)
print(response)