DeepSeek V4 Flash Released: Supports Inference Fine-tuning & Web Search, API Compatible with OpenAI Ecosystem

DeepSeek V4 Flash and V4 Pro updated; API compatible with OpenAI/Anthropic, vision-exp model enters experimental phase.

DeepSeek officially released the V4 Flash model on August 29, 2025, and simultaneously updated the V4 Pro variant. Key facts:

  • Release date: August 29, 2025 (document date)
  • New models: deepseek-v4-flash upgraded to DeepSeek-V4-Flash-0731; new experimental multimodal model deepseek-v4-flash-vision-exp added
  • Version update: deepseek-v4-pro upgraded to DeepSeek-V4-Pro-0813
  • Invocation method: Legacy interface names (deepseek-v4-flash / deepseek-v4-pro) are preserved and automatically point to the latest versions
  • Availability: API is live; developers can call immediately
  • Weight openness: Not mentioned as open-weight or open-source; explicitly a commercial API service

The API interface is designed to be compatible with both OpenAI and Anthropic API formats, enabling developers to reuse existing OpenAI SDKs or Anthropic-compatible toolchains with minimal changes.

Access Method and Development Integration

The documentation provides invocation examples in three mainstream programming languages:

  • curl command-line call: Supports toggle for stream (streaming) output
  • Python call: Requires installing the openai SDK; set base_url="https://api.deepseek.com" to point to the API endpoint
  • Node.js call: Requires installing the openai NPM package; configure baseURL to reuse OpenAI client code

All examples enable reasoning_effort: "high" and thinking: {"type": "enabled"}, indicating DeepSeek recommends enabling high reasoning intensity to fully leverage model capabilities. Notably, the newly added vision experimental model deepseek-v4-flash-vision-exp must be explicitly called by setting the model name and accepts image input, though input format and size limits are not specified in the documentation.

Agent Tool Ecosystem Integration

DeepSeek has also opened DeepSeek Harness to developer preview worldwide, targeting Agent builders globally. This component aims to simplify Agent toolchain integration; specific capabilities are detailed in the official guide.

More significantly, DeepSeek emphasizes that its API is already supported by numerous leading AI Agent and coding assistant tools—including Claude Code, GitHub Copilot, and OpenCode. Users need not write code: simply switch the backend model name in the tool configuration to deepseek-v4-flash or deepseek-v4-pro to start using V4 series capabilities instantly. This enables seamless “model plug-and-play” switching. Forrester analyst previously noted API interoperability is a key barrier to Model-as-a-Service (MaaS) adoption in 2025; DeepSeek’s approach significantly reduces integration friction for enterprise and individual developers.

Model NameVersion SuffixInput TypeReasoningNotes
deepseek-v4-flash0731TextEnabledDefault recommended model; reasoning intensity adjustable
deepseek-v4-pro0813TextEnabledHigh-quality output variant; reasoning intensity adjustable
deepseek-v4-flash-vision-expExperimentalText + ImageEnabledVision experimental model; no timeline for general availability mentioned

Practical Adoption Recommendations

  • Ready for immediate use:

    • Projects currently using OpenAI/GPT-4 APIs:只需 modifying base_url and API Key initialization; migration requires no code refactoring;
    • Teams already integrated with Claude Code or GitHub Copilot: Switch model name to deepseek-v4-flash or deepseek-v4-pro in tool settings for quick cost/performance validation;
    • Developers prioritizing high-reasoning tasks: Recommended to enable thinking: {"type": "enabled"} + reasoning_effort: "high", ideal for code generation, multi-step reasoning, etc.
  • Recommend waiting:

    • Production environments requiring stable multimodal input: deepseek-v4-flash-vision-exp is explicitly labeled experimental, not advised for live services yet;
    • Users with extreme demands on Chinese long-context or low-cost inference: Documentation does not disclose token pricing, context window length, or latency metrics; await official parameter补充 before final decisions.

Final Note

DeepSeek’s rapid V4 Flash update following its second funding round signals acceleration toward building a complete pipeline—from large model research to engineering deployment. The dual-compatibility strategy at the API layer—supporting both OpenAI and Anthropic formats—lowers developer migration barriers while avoiding ecological isolation. In today’s multi-model competitive landscape, this careful balance between “building underneath and watching the ecosystem overhead” may indeed be the critical survival strategy.