DeepSeek V4 Flash

deepseekdeepseek/deepseek-v4-flash

DeepSeek V4 Flash — fast, cost-efficient model with 1M context window. Supports reasoning, tool calling, and structured output.

DeepSeek V4 Preview announcement ↗Hugging Face — DeepSeek-V4-Flash ↗DeepSeek V4 technical report (PDF) ↗DeepSeek API pricing & thinking mode guide ↗

DeepSeek-V4-Flash is the fast, cost-optimized tier of DeepSeek's April 2026 V4 preview release. It is a 284B-total / 13B-active Mixture-of-Experts model that ships with a full 1M-token context window and Compressed Sparse Attention (CSA) architecture. Released alongside the larger V4-Pro (1.6T / 49B active), Flash targets production workloads that need frontier-level reasoning and agentic coding at dramatically lower per-token cost.

For TheRouter operators, V4-Flash is the default open-weight choice for high-volume reasoning, long-context RAG, and tool-using agents. It supports the same thinking-mode API as V4-Pro (non-thinking / high / max) and is fully OpenAI-compatible via api.therouter.ai/v1. DeepSeek reports that at 1M context, Flash achieves only 10% of V3.2's single-token FLOPs and 7% of KV cache size — the efficiency story that enables its aggressive pricing.

Best for

• High-volume production reasoning and agentic coding where Flash@max reasoning effort delivers near-Pro quality at ~3× lower output cost
• Long-context RAG and repository-scale analysis — 1M window with drastically reduced FLOPs and KV cache versus V3.2
• Tool-calling agents and structured JSON output at scale — OpenAI-compatible, supports tools + response_format, thinking mode returns reasoning_content
• Cost-sensitive workloads that previously routed to GPT-5.4 Mini or Gemini 3 Flash but now want higher reasoning quality at similar or lower price

Reach for something else if

• Maximum frontier intelligence — V4-Pro (or closed models) still leads on hard agentic benchmarks (Terminal-Bench, BrowseComp, SimpleQA-Verified) by 5–13 points
• Vision or multimodal input — V4-Flash is text-only; route image tasks to Claude Opus 4.7, Amazon Nova 2 Lite, or GPT-5.4
• Extremely latency-sensitive chat where even Flash's thinking overhead is unacceptable — non-thinking mode or smaller models may be preferable

How TheRouter serves this differently from the vendor

As the vendor operates it

DeepSeek's first-party service exposes V4-Flash as deepseek-v4-flash through OpenAI-format, Anthropic-format, and Responses API endpoints; Hugging Face also publishes the open weights for self-hosting.

On TheRouter

TheRouter serves the same text model under deepseek/deepseek-v4-flash on its OpenAI-compatible /v1 surface. Pricing, context length, output limit, and supported parameters are rendered from the live model registry rather than duplicated in this curated page.

api-docs.deepseek.com ↗

Context Length

Max Output

384K

Input Priceper 1M tokens

$0.162/ 1M tokens

Output Priceper 1M tokens

$0.648/ 1M tokens

Modalities

text→text

Pricing Breakdown

Type	Rate
Input	$0.162 / 1M tokens
Output	$0.648 / 1M tokens

Supported Parameters

temperaturetop_pfrequency_penaltypresence_penaltymax_tokensmax_completion_tokensstoptoolstool_choiceresponse_formatstreamreasoning

Specifications

Release date (preview)	2026-04-24api-docs.deepseek.com ↗	verified
Architecture	284B total / 13B active MoE; Compressed Sparse Attention (CSA) + DeepSeek Sparse Attention (DSA); 1M context nativeapi-docs.deepseek.com ↗	verified
Context window	1,000,000 tokensapi-docs.deepseek.com ↗	verified
Max output	384,000 tokensapi-docs.deepseek.com ↗	verified
Reasoning modes	Non-thinking, Thinking (high / max); reasoning_content returned separatelyapi-docs.deepseek.com ↗	verified
License — code	MIThuggingface.co ↗	verified
License — weights	DeepSeek Model License (commercial use permitted)huggingface.co ↗	verified
Open weights	Yes — available on Hugging Facehuggingface.co ↗	verified
Training tokens	32T+huggingface.co ↗	verified

Benchmarks

Benchmark	Score	Source
Terminal Bench 2.0 V4-Flash Max result in DeepSeek's comparison across modes; V4-Pro Max is listed at 67.9%.	56.9%%	huggingface.co ↗
LiveCodeBench Pass@1 V4-Pro Max: 93.5%. Gap of 1.9 pp.	91.6%%	huggingface.co ↗
MMLU-Pro V4-Pro Max: 87.5. Gap of 1.3 pp.	86.2	huggingface.co ↗
GPQA Diamond V4-Pro Max: 90.1. Gap of 2.0 pp.	88.1	huggingface.co ↗
SWE-bench Verified V4-Pro Max: 80.6. Gap of 1.6 pp.	79.0	huggingface.co ↗
Codeforces rating (Max) V4-Pro Max: 3,206 Elo. Gap of 154 Elo points.	3,052 Elo	huggingface.co ↗

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL

curl https://api.therouter.ai/v1/chat/completions   -H "Content-Type: application/json"   -H "Authorization: Bearer $THE_ROUTER_API_KEY"   -d '{
    "model": "deepseek/deepseek-v4-flash",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference →

Chat completion

Standard OpenAI-compatible chat endpoint. Add reasoning_effort and the thinking extra_body to enable chain-of-thought.

cURL

curl https://api.therouter.ai/v1/chat/completions \
  -H "Authorization: Bearer $THEROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek/deepseek-v4-flash",
    "messages": [{"role":"user","content":"Explain quantum computing in simple terms"}],
    "reasoning_effort": "high",
    "extra_body": {"thinking": {"type": "enabled"}}
  }'

More from deepseek

deepseek/deepseek-v4-proLarger sibling (1.6T / 49B active) — higher intelligence, 5-point AA Index lead, ~3× higher output price.deepseek/deepseek-v3.2Previous generation (671B / 37B active) — still excellent price/performance; V4-Flash supersedes on long-context efficiency and agentic tasks.

Similar models

Cross-provider sibling models

anthropic/claude-sonnet-4.6

Comparable reasoning + tool use at higher per-token cost; stronger on some agentic benchmarks.

openai/gpt-5.4-mini

News & changes

2026-04-24

DeepSeek releases V4 preview — 1M context at open-weight frontier pricing

DeepSeek announced V4-Pro and V4-Flash on April 24, 2026. Both models introduce Compressed Sparse Attention enabling 1M context at dramatically lower FLOPs and KV cache than V3.2. Flash (284B/13B) targets cost-sensitive production workloads; Pro (1.6T/49B) targets maximum intelligence. Weights are MIT-licensed and available on Hugging Face.

re-authored by TheRouterapi-docs.deepseek.com ↗

Part of

api-integration api-migration coding-agents cost-optimization model-deprecation Model lifecycle tracker openai-compatible pricing reliability responses-api routing

Guides

Frequently asked

How does V4-Flash compare to V4-Pro on real agentic workloads?

On LiveCodeBench, SWE-bench Verified, and MMLU-Pro the gap is 1.3–1.9 points. On harder agent benchmarks (Terminal-Bench, BrowseComp, SimpleQA) the gap widens to 5–13 points. For most production coding and tool-use agents, Flash@max is within acceptable range of Pro@high while costing ~3× less per output token.

re-authored by TheRoutercodersera.com ↗

Does V4-Flash support the same thinking mode API as V4-Pro?

Yes. Both models expose the same thinking toggle via extra_body and the same reasoning_effort levels (high / max). The reasoning_content field is returned at the same level as content and must be concatenated into subsequent turns when tool calls are involved.

api-docs.deepseek.com ↗

What changed in the July 2026 V4-Flash API update?

DeepSeek's pricing page now lists the served model version as DeepSeek-V4-Flash-0731, keeps the 1M context window and 384K maximum output, and marks Responses API support as available for V4-Flash only at retrieval time. DeepSeek also stated that legacy deepseek-chat and deepseek-reasoner routes would retire after July 24, 2026 and were already routing to V4-Flash before retirement.

api-docs.deepseek.com ↗

Fact ledger — every claim on this page traces here

source	URL	retrieved
Release date (preview)	api-docs.deepseek.com ↗	2026-05-25	verified
Architecture	api-docs.deepseek.com ↗	2026-05-25	verified
Context window	api-docs.deepseek.com ↗	2026-05-25	verified
Max output	api-docs.deepseek.com ↗	2026-05-25	verified
Reasoning modes	api-docs.deepseek.com ↗	2026-05-25	verified
License — code	huggingface.co ↗	2026-05-25	verified
License — weights	huggingface.co ↗	2026-05-25	verified
Open weights	huggingface.co ↗	2026-05-25	verified
Training tokens	huggingface.co ↗	2026-08-12	verified
Terminal Bench 2.0	huggingface.co ↗	2026-08-12	verified
LiveCodeBench Pass@1	huggingface.co ↗	2026-08-12	verified
MMLU-Pro	huggingface.co ↗	2026-08-12	verified
GPQA Diamond	huggingface.co ↗	2026-08-12	verified
SWE-bench Verified	huggingface.co ↗	2026-08-12	verified
Codeforces rating (Max)	huggingface.co ↗	2026-08-12	verified
DeepSeek releases V4 preview — 1M context at open-weight frontier pricing	api-docs.deepseek.com ↗	2026-05-25	verified
How does V4-Flash compare to V4-Pro on real agentic workloads?	codersera.com ↗	2026-05-25	to verify
Does V4-Flash support the same thinking mode API as V4-Pro?	api-docs.deepseek.com ↗	2026-05-25	to verify
What changed in the July 2026 V4-Flash API update?	api-docs.deepseek.com ↗	2026-08-12	to verify

DeepSeek V4 Flash

How TheRouter serves this differently from the vendor

Modalities

Pricing Breakdown

Supported Parameters

Specifications

Benchmarks

API Usage Examples

API guide

Chat completion

More from deepseek

Similar models

News & changes

DeepSeek releases V4 preview — 1M context at open-weight frontier pricing

Part of

Related news

Guides

Frequently asked