DeepSeek V4 Pro
V4 Pro is the flagship model ID deepseek-v4-pro. Use it when the job needs long context, careful reasoning, or reading images — not when you only need a cheap everyday reply.
Model IDs, rate limits and pricing follow the official API docs.
What V4 Pro is for
DeepSeek V4 Pro is a 671B MoE flagship with about 1 million tokens of context, full vision, and the strongest reasoning in the V4 line. Published-style speed on this site’s Features snapshot is about 45 tok/s. On the API the model ID is deepseek-v4-pro. Weights, where published, are MIT. In Claude Code mappings, Opus-class names usually land here.
It is not a new brand. It is the heavy tier next to V4 Flash (37B, ~128K, basic vision, ~112 tok/s). Same OpenAI-compatible endpoint: https://api.deepseek.com/v1 — you only change the model string. If you searched DeepSeek V4 Pro, you want the model that can hold a large repo, a long PDF, or a screenshot — and think before it answers. DSpark can sit on top; it is an accelerator, not a second ID.
When to pick V4 Pro
Reach for Pro when the task is hard, long, or visual:
- Long documents, multi-file codebases, or anything near the 1M context window.
- Deep reasoning: math, architecture reviews, hard bugs, agent planning.
- Vision: charts, UI screenshots, scanned pages — the full image path, not Flash’s basic lane.
- Claude Code / Harness jobs you would map to Opus-class work.
What V4 Pro is not
Do not put the flagship in the same bucket as Flash, tools, or an SLA:
- Not V4 Flash. Everyday chat, drafts and high-volume calls belong on deepseek-v4-flash.
- Not a third company or a Vision SKU. You still call deepseek-v4-pro; there is no deepseek-vision ID.
- Not DSpark. DSpark only accelerates whichever model you already called.
- Not a guarantee of ~45 tok/s. That figure is a site snapshot, not a contract.
At a glance
- Model ID
- deepseek-v4-pro
- Size
- 671B MoE
- Context
- ~1M tokens
- Speed
- ~45 tok/s
- Vision
- Full
- License
- MIT (open weights where published)
How to start
Web first, then lock the ID, then the API:
- Web app: open the DeepSeek chat UI and run one real Pro job — a long PDF, a multi-file repo question, or a screenshot — not a hello-world prompt.
- Lock the model string deepseek-v4-pro. Do not look for a generic deepseek-v4 or a vision-only ID.
- API: POST to https://api.deepseek.com/v1 with model deepseek-v4-pro. The endpoint is OpenAI-compatible; keys come from the official platform, not this site.
- Claude Code: map Opus-class names here. Haiku-class everyday work should stay on Flash.
Pro vs Flash vs the doors
Same DeepSeek API, different jobs:
| Layer | Role | When to open this page |
|---|---|---|
| V4 Pro | 671B MoE · ~1M · full vision · ~45 tok/s · deepseek-v4-pro | Long context, hard reasoning, screenshots |
| V4 Flash | 37B · ~128K · basic vision · ~112 tok/s · deepseek-v4-flash | Everyday chat, drafts, high-volume calls |
| Web app | Browser chat, no model string to type | Feel Pro on one real task before you lock the ID |
| DeepSeek API | OpenAI-compatible https://api.deepseek.com/v1 | You are shipping a product or agent |
| Claude Code | Opus-class names map to Pro | The agent job is flagship-hard |
Boundaries and traps
- Sending million-token jobs or real vision work to Flash, then blaming the family. Full vision and ~1M context belong on Pro.
- Passing deepseek-v4, deepseek-vision, or a retired name such as deepseek-chat. The live flagship ID is deepseek-v4-pro.
- Copying ~45 tok/s into an SLA. Measure on your traces; DSpark and batch size still move the number.
- Looking for an API key on this site. Keys are issued on the official platform only.
- Mapping every Claude Code subagent to Pro. Haiku-class volume work should use deepseek-v4-flash.
Related guides
DeepSeek V4
The V4 family: how to pick Pro vs Flash
DeepSeek V4 Flash
Faster and cheaper for everyday use
Vision
Read charts, UI shots, and scans
DeepSeek API docs
OpenAI-compatible API for your apps
Features
Pro vs Flash matrix, vision, and DSpark on this site.
Benchmarks
Published scores — not just marketing copy.
DeepSeek AI
Chat, code, and analyze documents
FAQ
Is V4 Pro a separate product from DeepSeek AI?
No — different layers. DeepSeek AI is the company and the Hugging Face org. V4 Pro is the flagship model you chat with or call via API, ID deepseek-v4-pro. This page does not issue keys; apply on the official platform, then hit https://api.deepseek.com/v1.
Pro or Flash?
Use Pro for long context, full vision and hard reasoning: 671B MoE, ~1M tokens, ~45 tok/s. Use Flash for speed and cost on everyday tasks: 37B, ~128K, ~112 tok/s. Same OpenAI-compatible API; you only change the model string.
Where do I get an API key?
From the official DeepSeek platform — not this site. Then call https://api.deepseek.com/v1 with model deepseek-v4-pro. Rate limits and prices live in the official API docs and can change.
What model ID do I pass in the API?
deepseek-v4-pro. There is no generic deepseek-v4 SKU and no deepseek-vision ID. Charts, long PDFs and Opus-class agent work all use this string. Treat names like deepseek-chat as retired.
How does Claude Code map to Pro?
Opus-class names usually land on deepseek-v4-pro. Point the agent at https://api.deepseek.com/v1 and set that model. Haiku-class everyday work should stay on Flash — do not bind every subagent to the flagship.
Can I run V4 Pro locally?
Only if you have the hardware the model card assumes. Pro is a 671B MoE; most laptops and single workstation GPUs should start with Flash instead. Pull weights from Hugging Face, not from this site. Self-host DSpark may need an extra enable step.
Does Pro include DSpark and Vision?
Full vision is a Pro capability; you still call deepseek-v4-pro. DSpark is speculative decoding around that ID, not a third model. On the official API it is on by default. Self-host runtimes may need their own flag.
Is ~45 tok/s a guarantee?
No. It is a Features-page snapshot, not an SLA. Real speed depends on prompt length, batching and whether DSpark is on. Read the official docs and measure before you put a number in a contract.