DeepSeek V4 Pro and Flash: The April Release Behind the Summer Cycle

DeepSeek released the open-weight V4 Preview family on April 24, 2026. That date is twelve days before this series' May 6 to August 6 core window. We include it as a clearly labeled contextual exception because a summer model survey that names DeepSeek but silently assigns an API migration or repository update as the launch date would be less accurate than preserving the real chronology.

The official release note introduced DeepSeek-V4-Pro at 1.6 trillion total and 49 billion active parameters, and V4-Flash at 284 billion total and 13 billion active. Both offer one-million-token context, thinking and non-thinking modes, API access, and downloadable MIT-licensed weights.

Why April 24 is the correct date

DeepSeek's API change log records V4-Pro and V4-Flash on April 24. Later model-card revisions, quantizations, and a July 24 retirement of the deepseek-chat and deepseek-reasoner aliases are not new foundation-model releases. During the transition those aliases pointed to V4-Flash non-thinking and thinking modes.

This distinction matters for market chronology and reproducibility. The announcement date answers when the family became available; a repository revision identifies the artifact tested. An audit record needs both, but publishedAt must not be moved forward to make an older release fit a newer window.

The core summer ledger therefore excludes V4 from its date count while this article supplies the DeepSeek context explicitly requested for the major-lab review.

Architecture and open artifacts

V4-Pro is a large mixture-of-experts model with 1.6T stored parameters and 49B activated per token. V4-Flash stores 284B and activates 13B. Both support one million tokens. DeepSeek says pretraining used more than 32 trillion tokens, followed by domain-specific SFT and GRPO reinforcement learning and then on-policy distillation into unified models.

The architecture combines Compressed Sparse Attention and Heavily Compressed Attention. DeepSeek reports that, at one-million-token context, Pro needs 27% of the single-token inference FLOPs and 10% of the KV cache required by V3.2. Manifold-Constrained Hyper-Connections target stable signal propagation, while the Muon optimizer targets convergence and training stability.

The official V4-Pro model card provides base and instruct weights, an encoding implementation, local-inference guidance, evaluation tables, and the MIT license. Open weights allow inspection and self-hosting, but a 1.6T checkpoint remains a multi-node infrastructure project.

First-party Pro Max benchmark table

DeepSeek reports the following for V4-Pro at Max reasoning effort:

These values come from DeepSeek's model card, including its harnesses, sampling, and Max budget. The same table includes competitor figures, but cross-provider rows should not be treated as a neutral leaderboard. Reasoning effort and agent tooling materially affect the result.

Pro, Flash, and reasoning effort

DeepSeek publishes a particularly useful within-family matrix. Flash moves from 71.2 on GPQA in non-thinking mode to 87.4 at High and 88.1 at Max; Pro moves from 72.9 to 89.1 and 90.1. On LiveCodeBench, Flash rises from 55.2 to 91.6, while Pro rises from 56.8 to 93.5.

Long context shows the same budget sensitivity. Flash MRCR at one million tokens is 37.5 non-thinking, 76.9 High, and 78.7 Max. Pro scores 44.7, 83.3, and 83.5. A one-million-token window without sufficient reasoning does not produce the headline score.

Flash approaches Pro on some reasoning tasks when given a larger budget, but Pro keeps a knowledge and hardest-agent advantage. Routing should compare cost per accepted task, not parameter count. Keep Flash as the high-throughput candidate and escalate to Pro only where its gain is measurable.

Agent and coding evidence

At Pro Max, SWE-bench Verified is 80.6%, SWE-bench Pro 55.4%, SWE Multilingual 76.2%, Terminal-Bench 2.0 67.9%, BrowseComp 83.4%, HLE with tools 48.2%, MCPAtlas 73.6%, and Toolathlon 51.8%. The spread is informative: strong web and repository results do not imply equally strong terminal or heterogeneous-tool completion.

DeepSeek's own comparison across modes shows agent scores improving with effort, but not always monotonically. MCPAtlas is 69.4, 74.2, and 73.6 for Pro non-think, High, and Max. More reasoning is not automatically better when tool interaction and latency are involved.

Local evaluation should include one neutral scaffold and the production scaffold, with identical repositories, tool schemas, and ceilings. The open-model operations guide covers revision pinning, dependency review, sandboxing, and staged rollout for downloadable checkpoints.

Million-token context is an engineering claim

The architecture's long-context efficiency is a major contribution. A 90% KV-cache reduction relative to V3.2 at one million tokens can change serving feasibility. Yet a percentage reduction from a very large baseline may still require substantial accelerator memory, networking, and cache management.

MRCR Pro Max at 83.5 and CorpusQA at 62.0 show capability and remaining error. Teams should not fill the context simply because it exists. Retrieval, summaries, provenance, and task-local working sets can improve accuracy and reduce cost.

Self-hosted validation should plot latency and memory from short prompts through 128K, 384K, and one million tokens. Test distractor density, evidence position, multi-turn tools, and cache eviction. The context engineering guide explains why a capacity limit and effective recall are different properties.

Thinking-mode integration details

The DeepSeek thinking-mode documentation supports toggling reasoning and selecting low, high, or max effort across OpenAI-, Anthropic-, and Responses-compatible interfaces. Thinking is enabled by default at high. The mapping is model-specific: a requested low effort maps to high on V4-Pro in the documented configuration.

When tools are used, reasoning_content must be passed back on subsequent subrequests or the API returns a 400 error. Temperature and top-p settings are ignored in thinking mode. These are not cosmetic SDK details; they can explain failed agents and non-reproducible benchmark runs.

DeepSeek recommends at least 384K context for Pro Max local operation and provides a dedicated encoding folder rather than a Jinja chat template. Pin that code with the weights. Loading custom model code should occur in an isolated environment after review.

Deployment, license, and migration

MIT licensing is permissive, and DeepSeek publishes mixed FP4/FP8 instruct checkpoints plus FP8 base models. Pro uses FP4 for MoE experts and FP8 for most other parameters. Operators still need storage, fast interconnect, expert parallelism, compatible kernels, checksum verification, and a documented quantization path.

The API supports both OpenAI Chat Completions and Anthropic-compatible calls without changing the DeepSeek base URL. Migration from legacy aliases should use the explicit deepseek-v4-pro or deepseek-v4-flash IDs. Alias retirement on July 24 is an operational deadline, not a release to republish as news.

Benchmark the hosted service before self-hosting, then compare exact prompts and accepted outcomes. Keep supply-chain manifests for every shard, tokenizer, encoding script, container, kernel, and inference engine.

Verdict and window status

DeepSeek V4 is an important open-weight family: it combines sparse 1.6T/49B and 284B/13B operating points, one-million-token context, strong coding and agent benchmarks, an explicit reasoning-effort matrix, and MIT artifacts. It belongs in any technical account of the models shaping summer 2026.

It does not belong in the May 6–August 6 count as an in-window launch. The correct date is April 24. Preserving that exception makes the series more complete without falsifying chronology. Teams evaluating V4 should begin with Flash and Pro High under a controlled cost-and-quality test, reserve Max for cases that justify it, and treat the open deployment as a serious distributed-systems program.

Source notes — reviewed 2026

• DeepSeek's V4 Preview release supplies the April 24 announcement, parameters, access, architecture summary, and API migration note.
• DeepSeek's official change log confirms the release date and distinguishes the later alias retirement.
• The DeepSeek V4-Pro model card and weights provide architecture, open artifacts, benchmark tables, license, and local deployment guidance.
• DeepSeek's thinking-mode guide documents effort mapping, tool-call state, and API behavior.
• The official V4 model collection records the Pro and Flash artifact family.