DeepSeek V4 officially released: millions of context open source models, fully upgraded agent capabilities

DeepSeek released V4 preview on April 24, 2026, with two models — V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active). Both models support 1M token context window, a first for an open-weight model at this scale. V4-Pro is immediately available via API at $1.74/M tokens input, with the model weights open-sourced on Hugging Face. The announcement garnered 1,092 points and 736 comments on Hacker News within hours, making it one of the most discussed AI releases of 2026.

Background and trigger events

Event Highlights:

• V4-Pro reaches the open source model SOTA on the Agent Coding benchmark, and the official claims to be better than Claude Opus 4.6 Max
• V4-Flash is an economical option with lower inference cost and better long context efficiency than Pro
• Introducing the DSA (Deeply Sparse Attention) architecture to control the cost of million-level context reasoning at the same level as short contexts
• Model weights have been open sourced in Hugging Face, and the API is available today

Key Impact

Technical Impact Table:

Industry Impact Table:

Adaptation suggestions

To Developers:

• V4-Pro is the preferred recommendation for Agent/coding workflow, Flash is not recommended for complex Agent tasks
• API is online, you can control the intensity of thinking through reasoning_effort parameter
• Open source weights can be deployed in Atlas Cloud, and self-hosted solutions can be compared with vLLM performance optimization

For users:

• DeepSeek web and APP have been simultaneously updated to the V4 series
• 1M context is suitable for processing long document analysis, code base level reasoning, multi-turn conversations and other scenarios

Points worthy of attention:

• Technical details and actual inference speed performance of DSA (Deeply Sparse Attention) architecture
• Validation of Agent capabilities by the open source community and third-party benchmark test results
• Actual measurement comparison of OpenClaw, Claude Code and other Agent tools

Industry reaction

Developer Community:

• The DeepSeek V4 post on Hacker News received 1,092 points and 736 comments within a few hours, becoming the most popular discussion of the day
• Developers are generally concerned about the actual availability of 1M context and the performance of the DSA architecture
• The community responded enthusiastically to the pricing of V4-Flash ($0.14/M tokens), believing that it will significantly lower the threshold for Agent applications

Competitors:

• OpenAI and Anthropic have yet to publicly respond to the release of DeepSeek V4
• Only a week ago (April 17), Anthropic just released Claude Design, and Claude Opus 4.7 was launched on April 16

Analyst Opinion:

"The 1M context is the real headline. V4 inference cost stays flat as tokens scale up vs V3.2 which shoots up. The architecture improvement is what makes this actually usable, not just a spec number." — HN Community Discussion

Tool entry embedding

This article naturally quotes the following tool entries (≥3 requirements) in technical details and industry analysis:

• Anthropic — cited in competitive landscape analysis
• OpenAI — mentioned in competitor response
• Claude Code — Mentioned in Ecological Integration and Agent Adaptation
• OpenClaw — the adapted Agent product mentioned in DeepSeek’s official announcement
• vLLM — mentioned in Technical Adaptation and Inference Optimization