DeepSeek V4 Flash 0731

deepseek · deepseek/deepseek-v4-flash-0731

Official DeepSeek V4 Flash release with enhanced agentic capabilities and integrated DSpark speculative decoding

≈667k Chinese characters of context · Step-by-step reasoning · Calls tools / agents · Structured JSON output · Open weights, self-hostable

Read a 200,000-Chinese-character book and write a 2,000-Chinese-character summary: about No public rate

An order-of-magnitude estimate at 1.5 tokens per Chinese character and the vendor's own rate — not a quote. Full method: the glossary on the AI Models page.

Specs

Context1,000,000(≈667k Chinese characters)
Max output384,000
Structured outputYes
Open weightsYes
Released2026-07-31

Source: models.dev snapshot 2026-08-29

Who serves it, at what price

ProviderIn /MOut /MCache read /M
Alibaba$0.2$0.4$0.04
Alibaba Token PlanBilled by planBilled by plan$0
Alibaba Token Plan (China)Billed by planBilled by plan$0
Cortecs$0.13$0.28$0.03
CrofAI$0.08$0.1$0.003
DigitalOcean$0.08$0.252$0.0252
EmpirioLabs AI$0.424$1.272$0.424
GreenPT$0.1596$0.399$0.0456
Charm Hyper$0.44$1.32$0.044
Requesty$0.14$0.28$0.07
Scaleway$0.468$0.936$0.0936
Venice AI$0.175$0.35$0.035

Rows tagged “first-party” are the model vendor's own pricing; untagged rows are resale channels and may differ.

When to choose it

We have not assessed this model. It comes from the full models.dev snapshot — the specs are the vendor's published figures, and a price is the vendor's own only on rows tagged first-party; the rest are resale channels. The fit is simply not something anyone here has written.

Leaderboard

Not yet on any leaderboard we track. That does not mean the model performs poorly — it means we have not taken this snapshot for it yet, and we do not publish a score we have not measured.