DeepSeek V4
V4 is the family name. Two API IDs sit under it: deepseek-v4-pro (flagship) and deepseek-v4-flash (fast lane). There is no third V4 model to call.
Model IDs, context limits, tok/s and prices follow the official API docs and can change. Speed figures here match this site’s Features / API snapshot (~45 tok/s Pro, ~112 tok/s Flash).
What “DeepSeek V4” means
People search DeepSeek V4 when they want the current generation — not V3.1-Terminus, not a new company. V4 is the series. The two billed models you actually call are deepseek-v4-pro (671B MoE, ~1M tokens, full vision, ~45 tok/s) and deepseek-v4-flash (37B, ~128K tokens, basic vision, ~112 tok/s).
Same OpenAI-compatible endpoint: https://api.deepseek.com/v1. You change the model string. Older names such as deepseek-chat are retired. DSpark is a speculative-decoding accelerator that can sit on both tiers — it is not a model ID. Full vision belongs on Pro.
How to choose a tier
Start from the task, then lock one ID:
- Long PDFs, big repos, screenshots, hard reasoning → V4 Pro (deepseek-v4-pro).
- Chat, drafts, extraction, subagents, cost control → V4 Flash (deepseek-v4-flash).
- Unsure? Try the web app on one real task, then lock the ID.
- Shipping a product? Call https://api.deepseek.com/v1 and swap the model string later.
What V4 is not
Do not treat the family name as a third SKU:
- Not a third model. There is no deepseek-v4 ID to call. You pick Pro or Flash.
- Not V3.1-Terminus. That was the previous generation. New work should use V4 IDs.
- Not DSpark. DSpark accelerates whichever V4 tier you already called.
- Not the web app, the API, or Hugging Face weights. Those are doors to the same family.
Published-style snapshot
- V4 Pro ID
- deepseek-v4-pro
- V4 Flash ID
- deepseek-v4-flash
- Pro size / context
- 671B MoE · ~1M tokens
- Flash size / context
- 37B · ~128K tokens
- Pro speed / vision
- ~45 tok/s · full vision
- Flash speed / vision
- ~112 tok/s · basic vision
- API
- https://api.deepseek.com/v1 (OpenAI-compatible)
- DSpark
- Accelerator on both — not a model
How to start
Web first, then lock an ID, then the API:
- Open the web app and run one real task — a PDF, a repo question, or a screenshot — not a hello-world prompt.
- Decide the tier from that run. Lock deepseek-v4-pro or deepseek-v4-flash. Do not look for a generic deepseek-v4 ID.
- For a product, call https://api.deepseek.com/v1 with that model string. The endpoint is OpenAI-compatible; you can swap the ID later if the job changes.
- Self-host only if data cannot leave the building: pull weights from Hugging Face deepseek-ai, then serve Flash or Pro yourself. DSpark may need an explicit enable locally.
Pro vs Flash vs the doors
The series is the brain; the doors are separate:
| Option | What it is | When to pick it |
|---|---|---|
| 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 the family before you lock an ID |
| DeepSeek API | OpenAI-compatible https://api.deepseek.com/v1 | You are shipping a product |
| Open weights | Hugging Face deepseek-ai checkpoints | Data must stay on your GPU |
Boundaries and traps
- Passing deepseek-v4 or a retired name such as deepseek-chat. The live IDs are deepseek-v4-pro and deepseek-v4-flash.
- Sending million-token jobs or real vision work to Flash. Full vision and ~1M context belong on Pro.
- Treating DSpark as a third V4 model. It only accelerates the ID you already chose.
- Copying ~45 / ~112 tok/s into an SLA. Those are site snapshot figures, not a contract.
- Confusing the web app with the API. Chat first; integrate after the ID is locked.
Related guides
FAQ
Is V4 a third model besides Pro and Flash?
No. V4 is the family name. The only IDs you call are deepseek-v4-pro and deepseek-v4-flash. There is no deepseek-v4 SKU. If a tutorial still says “call V4”, it means pick one of those two strings on https://api.deepseek.com/v1.
What about V3.1-Terminus?
That was the previous generation. New work should use V4 IDs. Do not mix Terminus prompts, mappings, or billing assumptions with Pro / Flash. If an old pipeline still sends a V3.1 name, migrate it before you treat latency or quality as a V4 result.
Where should I start?
Web app first: run one real task. Then lock deepseek-v4-pro or deepseek-v4-flash from that run. Then, if you are shipping a product, call https://api.deepseek.com/v1 with that ID. Agent shells (Harness, Deep Code, Claude Code) come after the prompt and the ID are stable.
Pro or Flash?
Pro is the 671B MoE flagship: ~1M tokens, full vision, ~45 tok/s. Use it for long PDFs, big repos, screenshots, and hard reasoning. Flash is the 37B fast lane: ~128K tokens, basic vision, ~112 tok/s. Use it for chat, drafts, extraction, and high-volume calls. Same API; you only change the model string.
Is DSpark a third V4 model?
No. DSpark is speculative decoding around whichever V4 tier you already called. On the official API it is on by default. You never pass a dspark model ID. Self-host runtimes may need their own enable flag. Acceleration does not add vision or 1M context to Flash.
Can I still use deepseek-chat?
Treat names like deepseek-chat as retired. New requests should send deepseek-v4-pro or deepseek-v4-flash. Compatibility shims, if any, are temporary and live in official change notes — do not build a new product on a legacy alias.
Does Flash do vision?
Full vision belongs on V4 Pro (deepseek-v4-pro). Flash lists basic vision and is the wrong default for charts, UI screenshots, or scanned pages. If the job is “read this image and reason”, lock Pro. If you only needed a cheap text reply, stay on Flash.