DeepSeek releases the DeepSeek-V4 preview with a 1M-token context

Background

The headline figure of the V4 preview was its context window. DeepSeek set the context at one million tokens, a length that lets a model take in very large documents or long conversations in a single pass ^[1]. A context window is the amount of text a model can consider at one time.

What happened

On 24 April 2026 DeepSeek released DeepSeek-V4 in preview. The models were open-sourced and went live the same day ^[1]. The release consisted of two mixture-of-experts models. DeepSeek-V4-Pro has 1.6 trillion total parameters, with 49 billion active per token. DeepSeek-V4-Flash has 284 billion total parameters, with 13 billion active ^[1].

Key details

The 1M-token context became the default across all official DeepSeek services at the time of the release ^[1]. Both models support thinking and non-thinking modes, so users and developers can choose between step-by-step reasoning and direct answers ^[1].

The architecture combines Compressed Sparse Attention and Heavily Compressed Attention, uses Manifold-Constrained Hyper-Connections for residual connections, and was trained with the Muon optimizer ^[2]. The pre-training data exceeded 32 trillion tokens. The model then went through supervised fine-tuning, reinforcement learning with GRPO, and on-policy distillation ^[2]. The released weights use FP4 for the MoE experts and FP8 for most other parameters, under the MIT License ^[2].

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].

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

Impact

The preview made a long-context, open-weight model available in the company's own app and API, and gave external tools such as Claude Code and OpenCode a route to the models ^[1].