Alibaba Cloud added deepseek-v4.1-flash-1017 to its Token Plan Personal Edition catalogue on October 20, replacing the earlier flash model in the list. The post says DeepSeek rebuilt the model with self-feedback-based reinforcement learning, and that agent tasks finish faster with shorter outputs on the way to an answer. Published specifications carry over: a 1M-token context window, a maximum output of 393,216 tokens, text and image input with text-only output, and an implicit context cache that bills cached input at a tenth of the fresh rate. Thinking is controlled by an enable_thinking flag that sits outside the OpenAI specification, so plain HTTP takes it at the top level of the JSON body; the chain of thought is billed as output tokens and can run up to 384K tokens ahead of the answer. Function calling, web search and the Responses API are supported. Structured output supports JSON Object but not JSON Schema. Prefix completion, batch calls and fine-tuning are not supported on this model.
Our takeThe shape of the API says more than the spec sheet: a 1M-context model without prefix completion, batch calls or fine-tuning is built for interactive agent loops and little else, and since thinking tokens are billed as output, the budget setting is a quality decision too. JSON Object without JSON Schema means typed output still needs client-side validation. The write-up credits the reinforcement-learning change for faster tasks and shorter outputs without separating the two.