DeepSeek V4 Flash

DeepSeek V4 Flash is the smaller, faster, and less expensive member of the DeepSeek V4 family. The official DeepSeek-V4-Flash-0731 API release entered public beta on July 31, 2026, replacing the earlier preview build. It has 284 billion total parameters, activates 13 billion per token through its Mixture-of-Experts routing, supports a 1 million token context window, and can produce up to 384K output tokens through DeepSeek's API.

For a normal user, Flash is the V4 model to choose when speed and price matter more than squeezing out the last few points of capability. For a developer, it is the high-throughput production tier: an open-weight, MIT-licensed model with tool calling, JSON output, OpenAI-compatible Responses API support, an Anthropic-compatible endpoint, and both thinking and non-thinking operation.

Key specifications

Specification	DeepSeek V4 Flash
Official API build	DeepSeek-V4-Flash-0731
API model ID	deepseek-v4-flash
Release status	Official release in public beta
Total / active parameters	284B / 13B
Architecture	Mixture-of-Experts with CSA + HCA hybrid attention
Context / maximum output	1M / 384K tokens
Reasoning modes	Non-think, Think High, Think Max
API features	JSON, tools, Responses API, Anthropic API, prefix completion, FIM
Weight precision	FP4 experts plus mostly FP8 remaining parameters
License	MIT

Open weights do not mean lightweight deployment. A 284B-parameter checkpoint still requires serious multi-GPU infrastructure, quantization, or a hosted inference provider. The 13B active figure describes sparse compute per token; it is not the amount of memory needed to store the whole model.

Official benchmark table

These are DeepSeek's reported instruct-model results. Flash Max uses a larger reasoning budget than High, so it should not be read as a free accuracy upgrade.

Benchmark	Flash Non-think	Flash High	Flash Max	Pro Max reference
MMLU-Pro	83.0	86.4	86.2	87.5
GPQA Diamond	71.2	87.4	88.1	90.1
LiveCodeBench	55.2	88.4	91.6	93.5
Codeforces rating	—	2816	3052	3206
IMOAnswerBench	41.9	85.1	88.4	89.8
MRCR 1M	37.5	76.9	78.7	83.5
Terminal Bench 2.0	49.1	56.6	56.9	67.9
SWE Verified	73.7	78.6	79.0	80.6
BrowseComp	—	53.5	73.2	83.4
Toolathlon	40.7	43.5	47.8	51.8

The strongest Flash result is not always Max: MMLU-Pro is marginally higher in High. This is a useful warning against assuming more test-time compute improves every task. DeepSeek's results also use its own prompts and evaluation harnesses; independent testing may differ, especially for agent benchmarks that depend heavily on tools and scaffolding.

Flash versus Pro

Flash and Pro share the same 1M context, 384K maximum output, API feature set, and three reasoning modes. The main differences are capacity, cost, and difficult-task reliability. Flash activates 13B parameters versus Pro's 49B, and its launch API rates are roughly one-third of Pro's cache-miss input and output prices.

Flash Max comes close to Pro Max on GPQA Diamond, LiveCodeBench, and IMOAnswerBench, making it a credible default for coding and reasoning workloads. Pro's advantage is clearer on factual knowledge, long-context retrieval, terminal work, browsing, and complex tool use. A sensible routing system can send routine requests to Flash and escalate only hard or failed tasks to Pro.

API price

As of August 14, 2026, the listed Flash rates per million tokens are $0.0028 for cache-hit input, $0.14 for cache-miss input, and $0.28 for output. DeepSeek has scheduled a new time-of-day system for August 16: off-peak rates of $0.007, $0.22, and $0.66, and peak rates of $0.014, $0.44, and $1.32 respectively. Because that change was announced but not yet effective when this entry was updated, check DeepSeek's live pricing page before estimating production cost.

When to use DeepSeek V4 Flash

Flash fits customer support, extraction, summarization, code generation, repository search, document analysis, tool-using assistants, and agent systems where many calls must fit inside a fixed budget. It is also the practical V4 model for experiments that need open weights but cannot justify Pro's 1.6T-parameter storage footprint.

Choose Pro instead when the job depends on obscure factual knowledge, sustained browser research, difficult terminal execution, competitive programming, or the highest possible success rate on a long multi-step trajectory. For sensitive workloads, open weights make self-hosting possible, but teams still need to audit the model, serving stack, prompts, tool permissions, and output handling.

Bottom line

DeepSeek V4 Flash is no longer the April preview alias. Flash 0731 is the official public-beta build and the economical V4 default: smaller than Pro, much cheaper to call, and still competitive on coding and reasoning when given a higher thinking budget.