DeepSeek put DeepSeek V4.1 Flash 1017 online this Saturday, an update to the lightweight model the company launched in September. The name carries the date because the Chinese lab adopted that naming pattern in earlier versions, like V4 Flash 0731 and V4 Pro 0813, but what stands out here is not the nickname, it is the performance jump: on coding and agent automation tasks, the model now beats Claude Opus 5 on several benchmarks, something the September version did not do.
What actually changed
According to the technical report published on Hugging Face, V4.1 Flash 1017 is a minor architectural update that delivers substantial gains in reasoning, coding, agentic tasks and vision, while also reducing execution time and response length. In practice that means faster and cheaper tasks, since API call cost grows with the length of the generated response. DeepSeek calls this a Pareto improvement, a way of saying the model got better and cheaper at the same time, without trading one for the other.
The numbers that matter
On DeepSWE v1.1, which measures the ability to solve real software engineering tasks, the model rose from 74.2% to 77%, beating Claude Opus 5, which sits at 74%, and GPT 5.6 Sol, at 73%. On Terminal Bench 4.0, which tests command line use by agents, the jump was from 31.2% to 42.3%. On Automation Bench and on Agents' Last Exam, two tests focused on agents that execute tasks on their own, V4.1 Flash 1017 also leads the comparison group DeepSeek chose, which includes Kimi K3, GLM 5.3, GPT 5.6 Sol and Opus 5. On Humanity's Last Exam, a hard general knowledge test, the model went from 36.8% to 42.6%.
Open weights and immediate access
The model is already live on DeepSeek's website and app, and on the API under the same deepseek-flash identifier as before, with no price change. The weights were published on Hugging Face and ModelScope under the MIT license, the most permissive in the market, allowing unrestricted commercial use. The company also opened its own web search and fetch APIs to any developer, charging $0.75 per thousand calls, with free use through its harness until October 31 at 16:00 UTC. It is a direct invitation for agent builders to test the tool for free before billing kicks in.
Why this bothers the competition
DeepSeek had already shaken the market in September, when the original V4.1 Flash launch, with aggressive pricing, sent shares of rivals like MiniMax, Z.ai and even Alibaba tumbling on the Hong Kong exchange. This update reinforces the same playbook: instead of launching a brand new model that is expensive to train from scratch, the company iterates quickly on an existing lightweight model and manages to stay competitive with far more expensive closed models, including Anthropic's own Opus 5, on specific agent tasks.
The limit still worth checking
The cited benchmarks come from DeepSeek's own technical report, with no independent replication published so far. The company also did not detail the real cost of running the full 763 billion parameter version outside its own infrastructure, and the model does not ship a standard Jinja chat template, requiring custom implementation from anyone integrating it through open source. Beating Opus 5 on specific agent benchmarks does not mean beating it on every task: on Chartography and ZeroBench, visual reasoning tests, Opus 5 still comes out ahead.
The MaxAssistant take
The pace is the most important fact here: DeepSeek ships a free, faster open weight update less than a month and a half after the original model, beating closed competitors on coding and automation tasks without raising the price. That pressures any provider charging a premium for an agent model, because it shows it is possible to compete in this niche with a relatively light model and open access to the weights. It is worth waiting for independent testing before accepting the numbers end to end, but the immediate availability, the unchanged price and the MIT license already change the calculation for anyone deciding today which model to use in a coding agent.
Sources
DeepSeek API Docs: DeepSeek-V4.1-Flash-1017, Wish-Driven Programming Made Real; Hugging Face: deepseek-ai/DeepSeek-V4.1-Flash-1017.
