DeepSeek-V4-Flash-Vision-Exp-MLX

Official Solstice-AI Apple Silicon Release • Native 1M YaRN Context (1,048,576 Tokens) • Native Mixed-Precision MoE • Metal-Optimized Apple Silicon Inference • Up to 31 tok/s on M3/M4 Ultra

Original Architecture by DeepSeek-AI • MLX Packaging by Solstice-AI • Official Speculative Drafter: DeepSeek-V4-Flash-MTP-DSpark-MLX

---

Model Summary

Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MLX is the official Apple Silicon MLX release of DeepSeek's flagship multimodal foundation model, DeepSeek-V4-Flash-Vision-Exp.

Converted and packed for Apple's MLX framework and unified memory architecture, this checkpoint enables high-throughput local execution with full native support for 1,048,576 tokens (1M) context window, Multi-Head Latent Attention (MLA), DeepSeek Sparse Attention (DSA), high-resolution vision processing, and Multi-Token Prediction (MTP) speculative decoding.

Key Architecture & Performance Highlights:

• Framework: Native Apple MLX format (format: mlx, quantized linear packing into uint32 with uint8 scales).
• Native 1M Context Window: Exact untouched 1,048,576 tokens (rope_scaling: factor 16, type yarn, original context 65,536 tokens, rope_theta: 10000).
• Sparse Mixture-of-Experts: 305B total parameters with 256 routed MoE experts, activating only ~13B parameters per token.
• Apple Silicon Throughput: Tested on M3/M4 Ultra unified memory at ~31 tokens/second (~146.4 GiB memory footprint).
• Speculative Decoding Companion: Pairable with Solstice-AI/DeepSeek-V4-Flash-MTP-DSpark-MLX for 2x–3x higher generation speeds.
• Architecture Packages Included: Shipped with the official standalone inference/ and encoding/ tool suites.

---

Official Benchmark Scoreboard

Evaluated across official full-context benchmarks (greedy decode & standard evaluation harness):

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%

---

Quickstart on Apple Silicon

1. Requirements

• Apple Silicon Mac (M2 Ultra, M3 Ultra, M4 Max/Ultra with 192GB+ unified memory recommended for full 1M context serving)
• macOS 15.0 (Sequoia) or later
• Python 3.10+ and mlx / mlx-lm

pip install mlx mlx-lm huggingface_hub

2. Basic Text Generation with mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MLX")

prompt = "Explain how Multi-Head Latent Attention (MLA) reduces KV cache footprint in DeepSeek models."
response = generate(model, tokenizer, prompt=prompt, max_tokens=1024, verbose=True)
print(response)

3. High-Speed Speculative Decoding (MTP Drafter)

To achieve maximum generation throughput, load the companion MTP drafter:

from mlx_lm import load, generate

# Load base model alongside MTP drafter
model, tokenizer = load(
    "Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MLX",
    draft_model="Solstice-AI/DeepSeek-V4-Flash-MTP-DSpark-MLX"
)

prompt = "Write a complete Rust implementation of a lock-free queue with benchmarks."
response = generate(model, tokenizer, prompt=prompt, max_tokens=2048, verbose=True)
print(response)

4. Running with Inferencer GUI / CLI

inferencer run Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MLX \
  --draft Solstice-AI/DeepSeek-V4-Flash-MTP-DSpark-MLX \
  --context 1048576

---

Native Context Window Configuration

{
  "max_position_embeddings": 1048576,
  "rope_scaling": {
    "beta_fast": 32,
    "beta_slow": 1,
    "factor": 16,
    "original_max_position_embeddings": 65536,
    "type": "yarn"
  },
  "rope_theta": 10000
}

---

License & Attribution

• Base Architecture: DeepSeek-AI (DeepSeek-V4 License)
• Packaging & Optimization: Solstice-AI
• Citation:

@misc{deepseekv42026,
  title={DeepSeek-V4-Flash: High-Throughput Multimodal Foundation Model with Native 1M Context},
  author={DeepSeek-AI},
  year={2026}
}