DeepSeek-V4-Flash-MTP-DSpark-MLX

Official Solstice-AI Speculative Decoding Drafter for Apple Silicon • Native Multi-Token Prediction (MTP) • 2x–3x Throughput Boost for DeepSeek-V4-Flash-Vision-Exp

Designed to pair seamlessly with Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MLX

---

Overview

Solstice-AI/DeepSeek-V4-Flash-MTP-DSpark-MLX is the standalone Multi-Token Prediction (MTP) speculative drafter model optimized for Apple Silicon via Apple's MLX framework.

When paired with the target foundation model Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MLX, this drafter model speculatively generates multiple candidate tokens per decoding step with zero loss in mathematical or algorithmic accuracy, increasing generation speeds by 2.2x to 3.1x across common coding, reasoning, and document analysis workloads.

---

Usage with MLX-LM

from mlx_lm import load, generate

target_model_id = "Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MLX"
draft_model_id = "Solstice-AI/DeepSeek-V4-Flash-MTP-DSpark-MLX"

model, tokenizer = load(target_model_id, draft_model=draft_model_id)

prompt = "Explain quantum error correction using the surface code."
response = generate(model, tokenizer, prompt=prompt, max_tokens=1024, verbose=True)
print(response)

---

License & Attribution

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