DeepSeek pushes its model further with a faster version

DeepSeek launched DeepSeek-V4.1-Flash-1017 on its web, mobile, and developer interfaces. The update promises better results in reasoning, code, and agents, in an increasingly tight competition with OpenAI and Anthropic.

DeepSeek has deployed DeepSeek-V4.1-Flash-1017 on its website, mobile app, and developer offerings, with the stated goal of accelerating task execution and reducing resource consumption. According to the technical documentation published on Hugging Face, this version fine-tunes the V4.1-Flash model on reasoning, code, autonomous agents, game development, and 3D modeling. The launch is significant as it increases pressure on OpenAI, Anthropic, and Chinese rivals like Moonshot AI and Z.ai, while reminding that the race is no longer just about raw model quality but also about cost and infrastructure.

A speed and accuracy-focused update

DeepSeek presents this iteration as a leaner and faster version of its main model. Benchmark results published on X show a score ranging from 36.8 to 42.6 on Humanity’s Last Exam, compared to 36.8 to 39.1 for the previous generation. This gap should be interpreted cautiously, but it suggests a measurable gain on a test known to be difficult, covering thousands of questions in mathematics, science, and humanities.

The context is tense. Chinese developers of open models are trying to close the gap with American proprietary systems, which are often more costly to use. DeepSeek has already shown that an aggressive pricing policy can disrupt public markets: during the launch of its lightweight V4.1-Flash version in early September, several regional groups such as MiniMax Group, Z.ai, and Alibaba Group Holding Ltd saw their shares drop. The message is clear, even without fanfare: in this segment, price is not an accounting detail, it is a strategic signal.

The downside of growth: demand strains infrastructure

DeepSeek’s rise is not without setbacks. The company experienced a five-hour server outage early in the week, amid recurring disruptions that analysts attribute to bottlenecks in computing capacity. The more usage grows, the more the infrastructure is stressed. And, apparently, it has not yet found its balance.

This technical fragility does not prevent the company from expanding its scope. According to Bloomberg, DeepSeek is participating in a multi-million-dollar funding round for Winverse AI, a startup specializing in robotics founded in 2026 by Hu Han, a former researcher at Tencent, Baidu, and Microsoft Research Asia. The deal would also include Meituan via Long-Z Investments and could value Winverse AI at around 400 million dollars. At this stage, the signal is mainly strategic: DeepSeek is no longer only looking at language models but also hardware, physical automation, and embodied AI.

Competition now plays out on multiple fronts

In this segment, the battle is no longer limited to announced performance. DeepSeek must contend with established American players as well as Chinese competitors advancing rapidly on open models. The launch of DeepSeek-V4.1-Flash-1017 thus fits into a sequence where each update serves both as a technical demonstration and a message to the market.

What changes, above all, is the overall logic. DeepSeek pushes a more efficient model but must simultaneously support growing usage, maintain service availability, and prepare growth drivers beyond software. The AI market still rewards high-performing models. It is also beginning to value those that handle load, control costs, and know how to broaden their playing field.

Key points

  • DeepSeek-V4.1-Flash-1017 is deployed on web, mobile, and developer tools.
  • Humanity’s Last Exam: 36.8 to 42.6, compared to 36.8 to 39.1 previously.
  • A server outage lasted five hours early in the week.
  • Bloomberg mentions a funding round for Winverse AI.
  • Winverse AI could be valued at 400 million dollars.
  • Hu Han founded Winverse AI in 2026.

By the numbers

  • 5 hours — server outage reported at DeepSeek early in the week.
  • 36.8 to 42.6 — DeepSeek’s score on Humanity’s Last Exam, according to X.
  • 36.8 to 39.1 — score of the previous version on the same benchmark.
  • “Tens of millions of dollars” — amount targeted for Winverse AI’s funding round, according to Bloomberg.
  • 400 million dollars — potential valuation of Winverse AI, according to Bloomberg.

To read