DeepSeek V4 Flash
V4 Flash is the fast tier. Model ID deepseek-v4-flash. Use it for chat, drafts, high-volume API calls, and subagents — not for million-token jobs or heavy vision.
Official IDs, rate limits and prices follow the DeepSeek API docs.
What V4 Flash is for
DeepSeek V4 Flash is the 37B everyday model: about 128K context, ~112 tok/s, lower price, and only basic vision. On the API the ID is deepseek-v4-flash. Same OpenAI-compatible endpoint as Pro: https://api.deepseek.com/v1. Claude Code mappings send Haiku-class (and often Sonnet-class) names here.
It is still V4, just the light lane. If Pro is the 671B MoE flagship (~1M context, full vision, ~45 tok/s), Flash is the one you call a hundred times a day without watching the bill. Do not expect million-token windows or Pro-grade charts from this ID. DSpark can accelerate Flash on the hosted API; it does not turn Flash into Pro.
When to pick V4 Flash
Default to Flash unless the task clearly needs Pro:
- Everyday chat, email drafts, short summaries, and UI copy.
- High-volume API: classification, extraction, template fills.
- Claude Code subagents and Haiku-class work.
- Trying an idea in the web app before you pay for Pro tokens.
What V4 Flash is not
Do not treat the fast lane as a smaller Pro:
- Not V4 Pro. There is no ~1M context and no full vision on this ID.
- Not a third product or a DSpark model. You still call deepseek-v4-flash.
- Not the right default for dense charts, UI shots or scanned-page layout.
- Not a guarantee of ~112 tok/s. That figure is a site snapshot, not a contract.
At a glance
- Model ID
- deepseek-v4-flash
- Size
- 37B
- Context
- ~128K tokens
- Speed
- ~112 tok/s
- Vision
- Basic
- Best for
- Speed and cost
How to start
Web first, then lock the ID, then the API:
- Web app: open the DeepSeek chat UI and run one everyday task — a draft, a short summary, or a cheap extraction — so you feel the speed.
- Lock the model string deepseek-v4-flash. Do not look for a generic deepseek-v4 ID.
- API: POST to https://api.deepseek.com/v1 with model deepseek-v4-flash. The endpoint is OpenAI-compatible; keys come from the official platform, not this site.
- Claude Code: map Haiku-class (and often Sonnet-class) names here. Opus-class hard work belongs on Pro.
Flash vs Pro vs the doors
Same API, pick the lane:
| Layer | Role | When to open this page |
|---|---|---|
| V4 Flash | 37B · ~128K · basic vision · ~112 tok/s · deepseek-v4-flash | Everyday chat, drafts, high-volume calls |
| V4 Pro | 671B MoE · ~1M · full vision · ~45 tok/s · deepseek-v4-pro | Long context, hard reasoning, screenshots |
| Web app | Browser chat, no model string to type | Feel the speed before you lock the ID |
| DeepSeek API | OpenAI-compatible https://api.deepseek.com/v1 | You are shipping a product or agent |
| Claude Code | Haiku-class names map to Flash | Subagents and cheap everyday agent work |
Boundaries and traps
- Sending a million-token PDF or a dense chart to Flash, then calling the model “blind.” That job is Pro.
- Passing deepseek-v4 or a retired name such as deepseek-chat. The live fast ID is deepseek-v4-flash.
- Treating DSpark as a reason to skip Flash. DSpark only makes the ID you already chose answer faster.
- Copying ~112 tok/s into an SLA. Measure on your traces; batch size still dominates.
- Looking for an API key on this site. Keys are issued on the official platform only.
Related guides
DeepSeek V4
The V4 family: how to pick Pro vs Flash
DeepSeek V4 Pro
1M context, deep reasoning and vision
Local deploy
Run Flash on your GPU with Ollama or vLLM
DeepSeek API docs
OpenAI-compatible API for your apps
Features
Pro vs Flash matrix, vision, and DSpark on this site.
DeepSeek AI
Chat, code, and analyze documents
FAQ
Is Flash a different product from Pro?
Same V4 family, two tiers. Flash is the 37B fast lane: ~128K tokens, basic vision, ~112 tok/s. Pro is the 671B MoE flagship: ~1M tokens, full vision, ~45 tok/s. You switch by changing the model string on https://api.deepseek.com/v1.
Can I use Flash in Claude Code or Harness?
Yes. Point the agent at the DeepSeek API and set the model to deepseek-v4-flash (or the Haiku mapping). Opus-class hard jobs should still use deepseek-v4-pro. Subagents and cheap volume work belong on Flash.
Will Flash read a huge PDF?
Not as well as Pro. About 128K is plenty for many docs; million-token jobs belong on V4 Pro. If the file is a scan whose layout matters, that is a Pro vision job, not a Flash text job.
What model ID do I pass in the API?
deepseek-v4-flash. There is no generic deepseek-v4 SKU. Keys come from the official platform; this site does not sell them. Rate limits and prices follow the official API docs.
Does Flash do vision?
Only at a basic level — fine for a simple “what is in this photo?” prompt. Charts, UI screenshots and scan layout belong on Pro (deepseek-v4-pro). If pixels drive the answer, do not gamble on Flash.
Does DSpark replace Flash when I need speed?
No. Flash is the cheaper, smaller model. DSpark is speculative decoding around whichever ID you already called. On the official API it is on by default. Acceleration does not add 1M context or full vision.
Can I run Flash locally?
Yes — Flash is the usual laptop / single-GPU pick. Pull weights from Hugging Face, then serve with Ollama, vLLM or SGLang. Ollama is a third-party runtime. Self-host DSpark may need an extra enable step.
Is ~112 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.