DeepSeek-V4.1-Flash-1017 is an updated build of DeepSeek V4.1 Flash, and the deepseek-flash API name already serves it, with the same calling method and price. If your agents call that name, or the retired V4 Flash names, the model under them changed on 17 October. Rerun your evaluations and check output length before trusting it.
What was released
DeepSeek-V4.1-Flash-1017 was announced by DeepSeek on 17 October 2026 in its API change log, with open weights published the same day in the DeepSeek-V4.1-Flash-1017 model card on Hugging Face. DeepSeek calls it a minor update to V4.1 Flash, which it released on 10 September 2026. The web and app versions of DeepSeek were moved to the new build too.
What changed and what stayed the same
DeepSeek states that the new build improves reasoning, coding, agentic and visual results while cutting task execution time and output length, so the same task should finish faster and cost less. Its own figures, all from the change log, include:
- Terminal-Bench 4.0: 42.3, up from 31.2 for the September V4.1 Flash, run in the Claude Code harness.
- DeepSWE v1.1: 77.0, up from 74.2.
- Automation-Bench: 60.5, up from 54.8.
These are the lab's numbers on its chosen harnesses, not a guarantee for your workload. DeepSeek also says it tuned the build for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and asks users of its own harness to move to version 0.2.1-rc.1 or later.
What did not change, according to the Models and Pricing page:
- Name and price: deepseek-flash, billed at the existing Flash rates, with off-peak rates at half the peak rate.
- Limits: a 1M-token context and up to 384K output tokens, with thinking mode on by default.
- Interfaces: JSON output, tool calls, an OpenAI-format base URL, a Responses API and an Anthropic-format base URL.
- Legacy names: deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but they point at the retired models and are now served by the 1017 build.
The weights stay under the MIT licence. The underlying V4.1 Flash design, set out in its base model card, is a multimodal mixture-of-experts model with 552B backbone parameters that activates 8B per token while reading input and 16B while generating, with a reasoning effort setting from 1 to 100.
Why a silent upgrade matters for agents
The practical point is that nobody on your team chose this upgrade. The pricing page lists only the 1017 build behind deepseek-flash, so a production agent calling it started using a different model on 17 October without a deploy on your side. Better scores on a lab's benchmark do not tell you that your prompts, tool schemas and output parsers behave the same way.
In our experience the regressions after a model swap are rarely about raw ability. They are about shape: a tool call with a different argument order, a JSON field that is now omitted, a shorter answer that drops a caveat your reviewers rely on. DeepSeek says the new build writes shorter outputs, which is good for cost but is exactly the kind of change that breaks a downstream check written against the old length or structure.
What to do this week
- Rerun your evaluation set on deepseek-flash now and compare with your last stored results, not with DeepSeek's tables. Our note on what AI agent evals catch covers what that set should contain.
- Check tool calls and JSON against your schemas, including the reasoning effort and thinking mode your agents send.
- Watch output tokens and latency on your dashboards for the days either side of 17 October, so you can see whether the promised cost and speed gain shows up in your traffic.
- Find calls to the old names. deepseek-v4-flash still works, but it now means the 1017 build, and the name no longer says what it runs.
- Record the model build in every trace, so the next silent change is visible. Our guide to versioning an AI agent for rollback explains how.
When not to rely on the hosted alias
If your agent needs a fixed model for audit or regulated work, a moving alias is the wrong dependency. The API offers no older Flash build to pin, so the options are to self-host the open weights at a version you control, as described in our guide to running an AI agent on a self-hosted model, or to keep a second provider tested as a fallback. Treat each provider change log as part of your release process, as our note on what to do when a model is deprecated suggests.
Frequently asked questions
What is DeepSeek-V4.1-Flash-1017?
It is an updated build of DeepSeek V4.1 Flash, announced by DeepSeek on 17 October 2026. DeepSeek describes it as a minor update with gains in reasoning, coding, agentic and visual tasks, plus shorter outputs and faster task completion. It replaces the September V4.1 Flash on the DeepSeek API, web and app, and its weights are on Hugging Face under the MIT licence.
Do I need to change my code to use DeepSeek V4.1 Flash-1017?
No. DeepSeek says the calling method and pricing are unchanged, and the deepseek-flash model name now serves the 1017 build automatically. Calls that still use deepseek-v4-flash or deepseek-v4-flash-vision-exp are also routed to it and billed at the Flash price. The work is in testing, not code: rerun your evaluations and check tool calls and output length.
Can I keep using the previous DeepSeek V4.1 Flash build on the API?
DeepSeek's pricing page lists only the 1017 build behind deepseek-flash, so there is no older Flash version to select on the hosted API. If you need a model that cannot change underneath you, run the open weights yourself at a fixed version, or keep a tested fallback provider ready while you validate the new build.
Is DeepSeek V4.1 Flash-1017 cheaper to run?
The per-token price is unchanged, but DeepSeek says the new build completes tasks with shorter outputs and in less time, which should lower the cost of each task. Off-peak rates are half the peak rates, with peak hours set on weekday UTC mornings. Measure output tokens per task in your own traffic before and after 17 October to confirm the saving.
