DeepSeek releases V4 series models: 1M context, API compatible with OpenAI/Anthropic, zero CUDA dependency

Core conclusion

DeepSeek officially released the V4 series models on April 24, 2026, including deepseek-v4-flash and deepseek-v4-pro. Both models support a 1M token context window and a maximum output token count of 384K. The V4 series is compatible with both OpenAI and Anthropic formats at the API level, and developers can switch between them without additional adaptation. It is worth noting that V4 runs entirely on Huawei’s Ascend chip platform and does not rely on any CUDA, indicating that China’s AI ecosystem has a complete independent technology stack.

Background and trigger events

Event information table

Event highlights

• Fact 1: DeepSeek releases V4 series, including Flash (lightweight and fast) and Pro (flagship performance) versions
• Fact 2: Both models support 1M token context and 384K max output tokens
• Fact 3: The API format is compatible with both OpenAI SDK and Anthropic SDK protocols
• Fact 4: Flash version is priced at $0.14/M input (cache miss)/$0.28/M output; Pro version is priced at $1.74/M input/$3.48/M output
• Fact 5: Provides Thinking Mode, Tool Calls, JSON Output, FIM Completion, Chat Prefix Completion and other functions
• Fact 6: Model weights have been open sourced under the MIT license at Hugging Face (2171 stars/24 hours)

Technical features

• The old model names deepseek-chat and deepseek-reasoner will be deprecated on 2026/07/24 and map to the non-thinking mode and thinking mode of V4 Flash respectively
• Supports Prompt Caching (context caching), cache hits can be reduced to $0.028/M (Flash) / $0.145/M (Pro)
• Zero CUDA dependency, completely based on Huawei Ascend chip training and inference

Key Impact

Technical Impact Table

Industry Impact Table

Industry reaction

• HN Community: It received 1322 points and 949 comments within 11 hours after release, and the discussion was more popular than the release of GPT-5.5 at the same time. The community is generally concerned about its technological breakthroughs with zero CUDA dependency and extremely low API pricing.
• Developer Feedback: jari_mustonen wrote on HN: The level of open source is top-notch in this field, the developer documentation is first-rate, the price is ridiculously low, and it provides cutting-edge model capabilities.
• Geotech Discussion: Many commentators pointed out that DeepSeek V4 running entirely on Huawei chips is an important milestone, marking that the maturity of China's AI technology stack has reached a level that can compete with the NVIDIA ecosystem.

Adaptation suggestions

To Developers:

• Projects that have been integrated with OpenAI or Anthropic SDK only need to modify the base_url to switch to DeepSeek V4
• V4 Flash is suitable for high-frequency, low-latency scenarios (chatting, code completion); V4 Pro is suitable for complex reasoning and in-depth analysis
• Pay attention to the deprecation deadline of old model names on 2026/07/24, migrate in advance

For users:

• DeepSeek provides a third high-quality model selection beyond OpenAI and Anthropic, especially suitable for cost-sensitive scenarios
• 1M context window has significant advantages for long document analysis, code warehouse understanding and other scenarios

Points worthy of attention:

• DeepSeek V4’s Flash/Pro dual-track pricing model may become an industry standard
• The open source strategy gives DeepSeek a unique advantage in enterprise private deployment
• The domestic full-stack solution of Huawei chip + DeepSeek model may give rise to a new AI application ecosystem

Action guidance

Want to learn more about practical experience in multi-model integration and API adaptation? Please check out:

• Tutorial: How to integrate OpenAI, Anthropic and DeepSeek at the same time to build a flexible AI workflow
• Case: A practical case of using multi-model API strategy to reduce AI business costs