New Model Released: V4.1 Flash Becomes the Primary Offering

DeepSeek has officially launched the DeepSeek V4.1 Flash model as its new flagship inference model. Key facts:

• Release Date: September 2026 (documentation dated September 11)
• New Model Name: deepseek-flash (recommended); legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp remain accepted but are retired
• Legacy Deprecation: V4 Pro (deepseek-v4-pro) will retire at 12:00 Beijing Time on September 14, 2026, with all requests automatically routed to V4.1 Flash
• Pricing: Legacy requests will be billed at Flash-tier pricing; V4 Pro will use Flash pricing after deprecation
• Availability: Compatible with OpenAI/Anthropic API format; supports both streaming and non-streaming modes

V4.1 Flash surpasses V4 Pro across performance, cost efficiency, response speed, and total time—a significant milestone for DeepSeek’s inference capabilities.

API Integration and Technical Details

V4.1 Flash adopts a request format compatible with OpenAI/Anthropic APIs. Existing SDKs require only model name updates for seamless access. Core parameters include model: "deepseek-flash", reasoning_effort: "high", and standard OpenAI-style messages structure. Notably, the model supports thinking: {"type": "enabled"} to activate DeepSeek’s proprietary reasoning chain mode.

Current SDK examples cover curl, Python, and Node.js, all requiring baseURL: "https://api.deepseek.com" and API key via environment variables. For Python, after initializing the client, developers call client.chat.completions.create() with extra_body={"thinking": {"type": "enabled"}} to enable advanced reasoning.

Surprising Detail: Although V4.1 Flash outperforms V4 Pro across metrics, DeepSeek clarifyV4.1 Pro has not yet been released—the current V4 Pro deprecation is a temporary measure until V4.1 Pro launches.

Model Capability Comparison

Note 1: Legacy Flash versions are officially retired; requests are handled by V4.1 Flash.

Developer Adoption Guidance

Users Who Should Act Immediately:

• Projects currently using deepseek-v4-flash, deepseek-v4-flash-vision-exp, or deepseek-v4-pro should update to deepseek-flash before September 14
• Teams relying on older SDKs or tools must verify support for thinking and reasoning_effort parameters

Users Who Should Wait:

• Developers with heavy reliance on V4 Pro’s reasoning capabilities may wait for the official V4.1 Pro release for a more stable long-term solution
• Temporary testing workloads can continue using legacy APIs, but remember post-September 14 billing will use Flash-tier rates

Final Thoughts

DeepSeek’s rapid iteration underscores the competitive race in inference models, where performance, cost, and latency form a three-dimensional battleground. The immediate retirement without extended transition signals strong confidence in model stability—and adds pressure on developers to adapt quickly.