V4.1 Pro appears in benchmark settings
"DeepSeek V4.1 Pro," a top-tier model under development by Chinese AI company DeepSeek, has appeared as an evaluation target in the settings of a benchmark that measures AI performance. LuminaBench, an X account that tracks AI models, posted about it on Nov. 1 (6:46 p.m. the same day Japan time). It said the model was included among the targets of "SearchQA," which measures the ability to answer questions using text gathered through search. No scores have been disclosed. The poster suggested that the arrival of Pro, trained using OpenAI's "GPT-6 Astra" as a model, was near. DeepSeek has made no announcement about the matter.
The same account had also posted on Sept. 27 (Sept. 28 Japan time) that a model name, "deepseek-v4.1-pro," had entered early access for some users, writing that "it shouldn't be far away now." More than a month later, it has not been released.
🚨 DeepSeek V4.1 Pro looks to be in early access testing The slug "deepseek-v4.1-pro" is in per account early access, deliberately hidden from the public model picker It shouldn't be far away now
DeepSeek announced its launch in September
DeepSeek itself announced the existence of V4.1 Pro on Sept. 10. On that day, the company released "V4.1 Flash" as the smallest model in a new architecture series. At the same time, it said it would phase out its previous top-tier model, "V4 Pro," and announced that from Sept. 14, requests to V4 Pro would be handled by V4.1 Flash. It explained that this arrangement would continue until the release of V4.1-Pro.
DeepSeek later revised the plan and has continued to offer V4 Pro at the same price after Sept. 14. Developer API pricing per 1 million tokens is $1.32 for input and $3.96 for output for V4 Pro, and $0.30 for input and $1.20 for output for V4.1 Flash (both peak-hour rates, with half price at other times). In DeepSeek's announcement, V4.1 Flash scored 74.2 on DeepSWE, which tests solving software development tasks, beating V4 Pro's 62.7, while on GPQA Diamond, which tests expert knowledge, V4.1 Flash scored 90.9 and V4 Pro 92.4. V4.1 Pro's release date, size and pricing have not been announced.
- DeepSeek releases V4.1 Flash and announces V4.1 Pro is comingOfficial
- Continued availability of V4 Pro noted on pricing pageOfficial
- Post says "deepseek-v4.1-pro" has entered early accessLeak
- Improved version of V4.1 Flash, "V4.1-Flash-1017," releasedOfficial
- The Information reports distillation allegations over the improved versionReport
- Post says V4.1 Pro has appeared in SearchQA evaluation settingsLeak
Distillation allegations over the smaller version
The improved version released just before has been reported to have been trained on outputs from OpenAI's models, and whether V4.1 Pro was built using the same method is a key question. The Information reported on Oct. 28 (Oct. 29 Japan time) that V4.1-Flash-1017, released on Oct. 17, and Moonshot AI's "Kimi K3.1 Pro" relied heavily on GPT-6 Astra's outputs for training. The method, known as "distillation," uses a strong model's answers to train another model, and OpenAI's terms of use generally prohibit using its outputs to develop competing models, according to IBTimes UK. DeepSeek, Moonshot AI and OpenAI did not immediately respond to Reuters' requests for comment.
OpenAI said on Sept. 30 that it had stopped an organized effort, ongoing since July, to extract its models' reasoning processes. Anthropic also said in October that new countermeasures had reduced the data that could be collected for distillation by about 90% to 99%.
