開發者平台
DeepSeek Retires Legacy API Model Names, Requiring Agents to Switch to V4-Flash or V4-Pro
DeepSeek retired the `deepseek-chat` and `deepseek-reasoner` compatibility names on July 24, meaning API requests that have not been migrated may now fail outright. The new interface separates model selection from reasoning mode, making the upgrade more involved than simply replacing a model string.
DeepSeek retired the `deepseek-chat` and `deepseek-reasoner` names from its API at 15:59 UTC on July 24. The official model identifiers are now `deepseek-v4-flash` and `deepseek-v4-pro`. The legacy names previously mapped to V4-Flash in non-thinking and thinking modes, respectively. Applications, model gateways, or agent frameworks that still hard-code these names may now receive runtime errors instead of being automatically routed to the new models.
The migration involves more than renaming models. The new API controls reasoning through `thinking.type` and `reasoning_effort`, and the official examples cover both the OpenAI-compatible `/chat/completions` API and an Anthropic-compatible endpoint. Engineering teams should verify that their SDKs preserve additional fields, streaming event handlers correctly process thinking content, and tool-call parsers support the DSML/XML format used by V4. Changing only an environment variable while retaining assumptions based on the old JSON format may still cause tool arguments or reasoning content to be lost.
V4-Flash is a mixture-of-experts (MoE) model with 284B total parameters and 13B active parameters per token, while V4-Pro has 1.6T total parameters and 49B active parameters per token. Both are listed as supporting a one-million-token context window. A technical overview from Hugging Face states that, at one million tokens, Flash's per-token FLOPs and KV cache are approximately 10% and 7% of V3.2's, respectively. However, these figures come from the official technical report and should not be interpreted as latency guarantees for any specific cloud endpoint.
Deployers should now rerun tool-calling, long-context, cache-hit-rate, and cost tests using real production traces, while adding alerts for unknown models and 4xx errors to their model gateways. V4-Pro and V4-Flash differ in quality, pricing, and reasoning behavior. Mechanically mapping every legacy alias to Pro may improve some tasks, but it will also affect latency and budgets.