The IT Home page is stamped 2026/10/20 15:57:37. Arena that day placed the already released DeepSeek-V4.1-Flash-1017 first on the open-model lists for Agent Arena and Code Arena: WebDev, and 11th overall. The Agent Arena row is named DeepSeek-V4.1-Flash-20261017 (Max), first among open models and 11th overall. “Strongest domestic model” is IT Home’s headline. Those four characters are not in the leaderboard sentences quoted above.

[1]
A closed stack of blank score cards
A closed stack of blank score cards, charcoal on top, muted blue edge, unmarked faces., AI-generated illustration, not a news photograph

The combined net improvement is 5.13%, from 7,498 real sessions. On the parts, confirmed task success is 11.54%, 8th overall; user feedback is 5.89%, 11th overall; Bash command reliability is 5.42%, 12th overall; steerability is 2.46%, 13th overall. The page calls all four first among open models. Tool hallucination is 0.48%. The page says this measures the model inventing tools it does not have, that a lower number means fewer of those cases, and that this row is 7th. The page does not call tool hallucination first among open models.

[1]

Cost is not one number. The earlier paragraph says the median cost per task is about $0.18. The later paragraph says the same row is about $0.18 to $0.19, and that different leaderboard pages differ slightly. The renminbi conversions in parentheses are not in this piece. The page also puts the model on the Pareto frontier of performance and task cost, meaning that under the metrics this evaluation used, it is one of the models that cannot be beaten on both performance and cost by the other selected models. That is a leaderboard classification, not the company’s own release note. When the model was revised, and what search costs, are not in the body of this page.

[1]

Both names are on this page. The announcement sentence uses DeepSeek-V4.1-Flash-1017. The leaderboard sentence uses DeepSeek-V4.1-Flash-20261017 (Max). The page does not say whether Max is a different tier, or whether 1017 and 20261017 are two ways of writing one release. Related links list the model release and a search interface. Numbers from those pages are not in the body here, and this edition does not import them. The Bash row is a reliability metric on the board, not a set of instructions.

[1]