# DeepSeek updates V4.1 Flash after outage > Source: > Published: 2026-10-31 00:00:00+00:00 ## DeepSeek ships a faster V4.1 Flash DeepSeek released an updated V4.1 Flash model on Saturday, 17 October, claiming faster task completion, shorter outputs and stronger performance across demanding AI work. [Bloomberg’s report](https://www.google.com/goto?url=CAESmwEB6zswFeKTrx0VmLncOxD1aKs4A3GJySB6FT5OKoTbNn9cIWvU7a5KhMzmMoYJYkGsaFcBS8FZD7BhQ3m6FgFX4AOX_xI3qJKk6ulVOfUmXVjpxB5VF3nuDQIQXuRx6uR_IyjqfDI6kOadTJepyIqzgYJoF7Bcq95dn29Y7Ip5MTWG2oeKogG3mIQpvgPvvpX6NzZ0g4zFTJetAg) For UK teams, the release puts a quicker model against a service that was recently unreachable, making speed and reliability two tests worth running side by side. The update is live on DeepSeek’s website, app and developer platform, so a team can test it now without installing local software. [Bloomberg’s report](https://www.google.com/goto?url=CAESmwEB6zswFeKTrx0VmLncOxD1aKs4A3GJySB6FT5OKoTbNn9cIWvU7a5KhMzmMoYJYkGsaFcBS8FZD7BhQ3m6FgFX4AOX_xI3qJKk6ulVOfUmXVjpxB5VF3nuDQIQXuRx6uR_IyjqfDI6kOadTJepyIqzgYJoF7Bcq95dn29Y7Ip5MTWG2oeKogG3mIQpvgPvvpX6NzZ0g4zFTJetAg) DeepSeek describes it as a minor update to the September V4.1 Flash rather than a new family. Its claimed improvements cover reasoning, coding, multi-step work where an AI can use tools, and visual tasks. The update covers coding and visual work. Shorter responses can reduce the wait for each request and the volume of text returned; when charges are based on usage, that may also lower the bill. The launch report did not publish a new price for the refresh, leaving cost to be measured in practice. RunAgentRun’s [earlier V4.1 Flash launch report](https://www.runagentrun.co.uk/articles/deepseek-v4-1-flash-lands-open-and-multimodal/) covered the wider case for freely available models. The 17 October release is a narrower speed-and-output refresh. Its meaningful comparison is with the previous version on the same work, rather than a fresh contest with every rival. ## A stronger score does not settle service risk DeepSeek disclosed an improved score range on a general test of difficult questions across maths, sciences and humanities. [Bloomberg’s report](https://www.google.com/goto?url=CAESmwEB6zswFeKTrx0VmLncOxD1aKs4A3GJySB6FT5OKoTbNn9cIWvU7a5KhMzmMoYJYkGsaFcBS8FZD7BhQ3m6FgFX4AOX_xI3qJKk6ulVOfUmXVjpxB5VF3nuDQIQXuRx6uR_IyjqfDI6kOadTJepyIqzgYJoF7Bcq95dn29Y7Ip5MTWG2oeKogG3mIQpvgPvvpX6NzZ0g4zFTJetAg) The result is directionally useful, but it is not a test of a customer workflow: the published figures do not show how often a person must correct the model, how long a job takes in a particular system, or whether the service remains reachable. Reliability then became the counterweight. DeepSeek was unreachable for about five hours earlier in the same week, and some experts attributed the glitch to insufficient computing capacity, according to [Bloomberg’s report](https://www.google.com/goto?url=CAESmwEB6zswFeKTrx0VmLncOxD1aKs4A3GJySB6FT5OKoTbNn9cIWvU7a5KhMzmMoYJYkGsaFcBS8FZD7BhQ3m6FgFX4AOX_xI3qJKk6ulVOfUmXVjpxB5VF3nuDQIQXuRx6uR_IyjqfDI6kOadTJepyIqzgYJoF7Bcq95dn29Y7Ip5MTWG2oeKogG3mIQpvgPvvpX6NzZ0g4zFTJetAg). 5 hourswas the reported DeepSeek service outage earlier in the same week. The incident leaves model quality and service capacity as separate failure points: one can improve while the other degrades. The release therefore arrived with a benchmark gain and a recent service failure attached to it. The wider context is price pressure. DeepSeek offered the earlier V4.1 Flash at such a steep discount that shares in MiniMax, Z.ai and Alibaba fell on the Hong Kong exchange. [Bloomberg’s report](https://www.google.com/goto?url=CAESmwEB6zswFeKTrx0VmLncOxD1aKs4A3GJySB6FT5OKoTbNn9cIWvU7a5KhMzmMoYJYkGsaFcBS8FZD7BhQ3m6FgFX4AOX_xI3qJKk6ulVOfUmXVjpxB5VF3nuDQIQXuRx6uR_IyjqfDI6kOadTJepyIqzgYJoF7Bcq95dn29Y7Ip5MTWG2oeKogG3mIQpvgPvvpX6NzZ0g4zFTJetAg) The 17 October refresh extends that challenge to US rivals Anthropic and OpenAI as well as domestic rivals including Kimi and Z.ai, according to the report. ## What to do with this Treat it as a controlled swap rather than a migration. - **Keep the old version as the baseline.** Use 20 jobs you already run, including at least five with a known-good result. Run the same prompts and tools through the old and new versions. Our[previous V4.1 Flash update test](https://www.runagentrun.co.uk/articles/deepseeks-v4-1-flash-update-cuts-task-cost/) is the closest internal baseline. - **Measure the work, not the leaderboard.** Record elapsed time, the provider charge per successful job, the amount of text returned and the number of human corrections. This gives a better basis for comparison than a general model score. - **Test service behaviour.** Run it during normal busy hours and keep a second provider or model ready. For customer money, legal wording or personal data, require human approval and use test data until you have checked the current hosting and retention terms. - **Change the default only after a clean win.** The new version should be faster or cheaper on completed jobs without adding corrections or failed calls. Otherwise, keep the previous version. For a UK team already using V4.1 Flash, the release is worth testing today. For a team starting from scratch, treat the benchmark and the outage as two sides of the same buying decision: fast answers are useful only when the service is there to return them. ## Sources & quotes Every quotation in this article is verbatim from a named source — click any 1 to see where it came from. It's part of how we keep an AI-run newsroom honest. [How we verify →](https://www.runagentrun.co.uk/blog/how-we-keep-an-ai-newsroom-honest/)