DeepSeek-V4 (V4-Pro and V4-Flash)

What it is

DeepSeek-V4 is a family of large language models from DeepSeek. It was released in preview on 24 April 2026, and both of its models were open-sourced and made available the same day ^[1]. The family has two mixture-of-experts (MoE) models. DeepSeek-V4-Pro has 1.6 trillion total parameters, of which 49 billion are active, and DeepSeek-V4-Flash has 284 billion total parameters, of which 13 billion are active ^[1].

A mixture-of-experts model holds many specialized sub-networks but activates only some of them for each token. Total parameters describe the full size of the model, while active parameters describe how much computation a single token uses.

Context and everyday use

Both models have a context window of one million tokens, and that length became the default across all of DeepSeek's official services ^[1]. Both models support thinking and non-thinking modes ^[1].

Architecture

DeepSeek describes V4 as a hybrid attention design that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) ^[2]. The model also uses Manifold-Constrained Hyper-Connections (mHC) for its residual connections, which are the shortcut paths that carry information from one layer to the next, and it was trained with the Muon optimizer ^[2].

The efficiency figures are the most specific. At one million tokens, DeepSeek says V4-Pro needs 27 percent of the single-token inference FLOPs and 10 percent of the KV cache of DeepSeek-V3.2 ^[2]. FLOPs measure the arithmetic work, and the KV cache stores the attention data a model reuses while it generates text, so both figures indicate how much work and memory long inputs cost.

Training

The model was pre-trained on more than 32 trillion tokens. Post-training then used supervised fine-tuning (SFT), reinforcement learning based on GRPO, and on-policy distillation ^[2]. The released weights use FP4 precision for the MoE experts and FP8 for most other parameters, and they are published under the MIT License ^[2].

Reported results

For V4-Pro at maximum reasoning effort, DeepSeek reported 80.6 on SWE-bench Verified, 67.9 on Terminal-Bench 2.0, 90.1 on GPQA Diamond, 37.7 on HLE, 93.5 on LiveCodeBench, a Codeforces rating of 3206 and 83.4 on BrowseComp ^[2]. DeepSeek also said that V4-Pro led open models on world knowledge, behind only Gemini-3.1-Pro, and that V4-Flash's reasoning came close to V4-Pro at lower prices ^[1].

API and integrations

The API supports both the OpenAI ChatCompletions format and the Anthropic API format, and V4 has integrations with Claude Code, OpenClaw and OpenCode ^[1]. The older model names deepseek-chat and deepseek-reasoner were retired after 24 July 2026 ^[1].

Timeline after the preview

V4-Flash became official on 31 July 2026 ^[6][3]. V4-Pro reached general availability on 13 August 2026 ^[3]. On 21 August 2026 an experimental multimodal model, DeepSeek-V4-Flash-Vision-Exp, accepting mixed text and image input, followed ^[4]. Wikipedia reports that V4 was adopted by Huawei and Cambricon Technologies ^[6].

On 10 September 2026 DeepSeek retired V4-Flash in favour of DeepSeek-V4.1-Flash. From 14 September, requests for V4-Pro were routed to V4.1-Flash until a V4.1-Pro model is launched ^[5].

Position in development

The combination that defines the release is a 1M-token context, a mixture-of-experts design and MIT-licensed weights ^[1][2].@@ZH_SUMMARY