DeepSeek Launches V4 Preview With 1M Context and Agentic Coding Focus

DeepSeekDeepSeek

· Updated

DeepSeek released the preview and open weights for DeepSeek-V4, a Mixture-of-Experts model family with a 1.6-trillion-parameter flagship and a 1M-token context window as the default. By introducing sparse attention and dual reasoning modes, the release delivers frontier-level agentic performance at lower compute costs.

DeepSeek, a Chinese research lab known for high-efficiency models, released the preview and open weights for DeepSeek-V4. The family includes DeepSeek-V4-Pro, a 1.6-trillion parameter Mixture-of-Experts (MoE) model (activating only some parameters per token), and a smaller DeepSeek-V4-Flash variant. Both models support a 1M-token context window as standard.
Pro parameters
1.6T total / 49B active
Flash parameters
284B total / 13B active
Context window
1M tokens
Reasoning modes
Thinking and Non-Thinking
API compatibility
OpenAI and Anthropic
Legacy retirement date
July 24, 2026

This release mirrors the pattern of preview MoE models but pushes context efficiency further through DeepSeek Sparse Attention. By compressing tokens, DeepSeek is commoditizing 1M-token memory, matching optimized inference paths in standard serving frameworks. It also introduces a configurable thinking mode for deep reasoning.

Access both models via the DeepSeek API, which supports OpenAI and Anthropic-compatible formats. The models are optimized for agentic coding and integrated with Claude Code, matching the support found in open-source agent platforms. Note that legacy deepseek-chat and deepseek-reasoner endpoints will be retired on July 24, 2026.

View the full update on api-docs.deepseek.com
DeepSeek
DeepSeek
@deepseek_ai
X

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length. 🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models. 🔹 DeepSeek-V4-Flash: 284B total / 13B active params. https://t.co/n1AgwMIymu

2.1kretweets10klikes
View on X

Still wondering? A few quick answers below.

DeepSeek-V4 is a family of open-weight Mixture-of-Experts models designed for high-performance reasoning and agentic tasks. It includes a 1.6-trillion parameter Pro version and a 284-billion parameter Flash version. Both models feature a 1-million token context window as the default standard, allowing the AI to process massive amounts of information in a single request.

The model uses a novel architectural innovation called DeepSeek Sparse Attention, or DSA, combined with token-wise compression. These techniques significantly reduce the computational and memory costs typically associated with long-context processing. This efficiency allows the 1-million token window to be the standard default across all official services without the usual performance or cost penalties.

The DeepSeek-V4 API supports both Thinking and Non-Thinking modes. Thinking mode allows the model to generate internal reasoning tokens to work through complex math, coding, or logic problems before providing a final answer. Non-Thinking mode provides faster response times for simpler tasks where deep deliberation is not required, giving users control over speed and accuracy.

DeepSeek-V4 is an open-weight model, meaning its parameters are publicly available for download. Developers can also access it via the official API, which supports OpenAI ChatCompletions and Anthropic API formats. Users simply need to update their model name to deepseek-v4-pro or deepseek-v4-flash while keeping their existing base URL and integration settings.

DeepSeek will officially retire the deepseek-chat and deepseek-reasoner model endpoints on July 24, 2026, at 15:59 UTC. Until that date, requests to these legacy models are being automatically routed to the new DeepSeek-V4-Flash architecture in either non-thinking or thinking modes to ensure a smooth transition for existing API users.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Share this update