DeepSeek reroutes its Flash API to a new MIT-licensed model

Developers who call DeepSeek's API are already running a different model without touching their code. According to DeepSeek's API documentation, the name deepseek-flash is now served by DeepSeek-V4.1-Flash-1017. The old aliases, deepseek-v4-flash and its experimental vision variant, still respond, but no model of their own sits behind them anymore: calls are forwarded to the new one and billed at the Flash rate.

The model's weights, the trained parameters needed to run it, are on the Hugging Face model card under an MIT license for both the repository and the weights, which allows commercial use. The card describes it as multimodal (it takes images and text and returns text), with 763 billion parameters and a 1-million-token context window, the amount of text it can read at once.

DeepSeek calls it a "minor update" to V4.1-Flash that, it says, runs tasks faster and answers more briefly, at lower cost. These are the results the company itself publishes:

All of these are DeepSeek's own figures, and no independent evaluation has been found. Other tables on the card pit it against rivals such as Opus 5, GPT-5.6 Sol, GLM-5.3 and K3, with mixed results: sometimes the rival wins.

On pricing, DeepSeek's rates page sets deepseek-flash at 0.30 dollars per million input tokens without cache and 1.20 dollars for output at peak hours; off-peak, half that. Peak hours run from 01:00 to 04:00 and from 06:00 to 10:00 UTC on weekdays.

For end users in Mexico there is no direct change, and no specific price or availability has been confirmed. The card gives no exact date: the 1017 in the name points to October 17, which the company has not confirmed.