DeepSeek V4 Model Card

DeepSeek V4 is DeepSeek's fourth-generation open-weight language-model family. The family first appeared in preview on April 24, 2026, but its two production-line variants have since received named releases: DeepSeek V4 Flash 0731 entered public beta on July 31, and DeepSeek V4 Pro 0813 launched across DeepSeek's app, website, and API on August 13. Calling the whole family a preview is therefore no longer accurate.

Both models are Mixture-of-Experts (MoE) systems with a 1 million token context window and up to 384K output tokens through DeepSeek's API. They support non-thinking and thinking operation, JSON output, tool calls, the Responses API, an Anthropic-compatible API, and beta prefix and fill-in-the-middle completion. Their open weights are released under the MIT License.

DeepSeek V4 models

Model	Official build	Release status	Total / active parameters	Context	Best fit
DeepSeek V4 Flash	DeepSeek-V4-Flash-0731	Official API release in public beta, July 31, 2026	284B / 13B	1M	Lower-cost, high-throughput agents and routine coding
DeepSeek V4 Pro	DeepSeek-V4-Pro-0813	General release on app, web, and API, August 13, 2026	1.6T / 49B	1M	Hard reasoning, coding, knowledge, and complex agent workflows

The active-parameter number is the better clue to per-token compute than the headline total. Flash stores 284B parameters but routes each token through about 13B; Pro stores 1.6T and activates about 49B. Pro is the higher-capability option, while Flash is designed to make the V4 architecture practical at much higher request volumes.

Architecture and training

DeepSeek V4 combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) for efficient long-context processing. DeepSeek reports that at a 1M-token context, V4 Pro requires 27% of DeepSeek V3.2's single-token inference FLOPs and 10% of its KV cache. The family also uses Manifold-Constrained Hyper-Connections (mHC), intended to stabilize signal propagation, and the Muon optimizer.

DeepSeek says the models were pretrained on more than 32 trillion tokens. Post-training combines supervised fine-tuning, GRPO-based reinforcement learning, specialist training, and on-policy distillation into unified instruct models. Both releases offer Non-think, Think High, and Think Max modes, allowing developers to trade latency and token use for stronger reasoning.

Official benchmark comparison

The following selected results come from DeepSeek's model card. They compare the released instruct models under the vendor's own evaluation settings; they are not independent tests.

Benchmark	V4 Flash High	V4 Flash Max	V4 Pro High	V4 Pro Max
MMLU-Pro	86.4	86.2	87.1	87.5
GPQA Diamond	87.4	88.1	89.1	90.1
LiveCodeBench	88.4	91.6	89.8	93.5
Codeforces rating	2816	3052	2919	3206
MRCR 1M	76.9	78.7	83.3	83.5
Terminal Bench 2.0	56.6	56.9	63.3	67.9
SWE Verified	78.6	79.0	79.4	80.6
BrowseComp	53.5	73.2	80.4	83.4
Toolathlon	43.5	47.8	49.0	51.8

The table shows the practical split. Flash Max can approach Pro on several reasoning and coding tests, but Pro retains a larger advantage on knowledge-heavy work, 1M-context retrieval, browsing, and difficult agentic tasks. Benchmark scores also depend on reasoning budgets, prompts, scaffolding, and tool harnesses, so buyers should reproduce representative workloads before choosing a tier.

API pricing and availability

As of August 14, 2026, DeepSeek lists Flash at $0.0028 per million cache-hit input tokens, $0.14 per million cache-miss input tokens, and $0.28 per million output tokens. Pro costs $0.003625, $0.435, and $0.87 respectively. DeepSeek has announced new peak and off-peak pricing effective August 16, so production budgets should use the live pricing page rather than treating these launch rates as permanent.

The API model IDs remain deepseek-v4-flash and deepseek-v4-pro. Developers who need stable, reproducible self-hosted artifacts can use the released Flash and Pro weights, while API users receive the dated 0731 and 0813 builds behind those names.

Which DeepSeek V4 model should you use?

Use Flash when cost, latency, or throughput dominates and the task can tolerate a modest capability tradeoff. Use Pro for difficult software engineering, competitive programming, research, long-context synthesis, browsing, and multi-step agents where a failed trajectory costs more than the model call. For either model, test High before Max: Max can improve hard-task results but usually increases latency and output-token consumption.

Bottom line

DeepSeek V4 is no longer best described as an April preview. It is a released two-tier family: Flash 0731 for efficient production inference and Pro 0813 for DeepSeek's strongest open-weight reasoning and agentic performance.