DeepSeek V4 Pro DSpark
deepseek-ai/DeepSeek-V4-Pro-DSpark
published Jun 2026 · updated Jun 2026
DeepSeek V4 Pro DSpark is a text-generation model that is a Mixture-of-Experts language model with speculative decoding for efficient inference, supporting million-token contexts.
specs
| Task | Text Generation |
| Architecture | Mixture-of-Experts (MoE) with Hybrid Attention (CSA + HCA), Manifold-Constrained Hyper-Connections |
| Parameters (Total) | 1.6T |
| Parameters (Activated) | 49B |
| Context Length | 1M tokens |
| License | MIT |
best for
- ·Long-context understanding and reasoning up to 1M tokens
- ·Code generation and software engineering tasks
- ·Complex multi-step reasoning and problem-solving
- ·Agentic workflows and tool use
FAQ
It excels at long-context tasks (up to 1M tokens), coding, complex reasoning, and agentic applications, with speculative decoding for faster inference.
DSpark is a speculative decoding module that increases throughput by generating multiple candidate tokens per step, reducing latency for the same checkpoint.
The model is released under the MIT License, allowing free use, modification, and distribution.
It uses an OpenAI-compatible chat format; see the model's encoding folder for message-to-token conversion and parsing.
Use the gigarouter OpenAI-compatible endpoint with your API key, setting the model name to deepseek-ai/DeepSeek-V4-Pro-DSpark.
We're benchmarking and onboarding DeepSeek V4 Pro DSpark as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.