Foreign media say DeepSeek and Kimi's new-model training made extensive use of GPT-6 Astra outputs

AT A GLANCE

The report states that DeepSeek the new version may make extensive use of Astra's outputs for distillation, but this has not yet been confirmed by official or independent evidence.

Article

DeepSeek Why has the new version suddenly become stronger? Foreign media reveals the inside story of Astra distillation, as Liang Wenfeng yields to competitive pressure

Dongcha Beating AI breaking news,DeepSeek the new model’s capabilities have surged, and Astra may genuinely deserve some of the credit.

Citing multiple people with direct knowledge of the matter,DeepSeek October 17 updated V4.1-Flash-1017, as well as Moonshot AI’s newly released Kimi K3.1 Pro, made extensive use of outputs from OpenAI’s GPT-6 Astra during training for distillation.DeepSeek The new version’s performance on browser-operation tasks is particularly impressive.Browser Use’s 180 real-world web task tests, its score rose from 70.9 in the previous version to 75.0, approaching Astra’s 77.9. Yet the model-call cost per question was only 0.22 USD, compared with Astra’s 10.77 USD, a difference of nearly 49 times.

Previously, Claude had consistently been the primary distillation target for domestic AI companies. But Anthropic has recently tightened its safeguards continuously for 2 months, greatly increasing the difficulty of extracting the model’s reasoning process, so domestic laboratories turned their attention to Astra.DeepSeek Liang Wenfeng, the founder of V4, was originally unwilling to rely on distillation in GLM development. But after seeing domestic competitors such as Zhipu AI make progress with the help of distillation, he ultimately yielded to competitive pressure.

Original link https://m.theblockbeats.info/flash/374297

Key points

The Information cited multiple people with direct knowledge as saying that DeepSeek V4.1-Flash-1017 and Moonshot AI's Kimi K3.1 Pro made extensive use of OpenAI GPT-6 Astra outputs for distillation during training.
DeepSeek In Browser Use its 180 real-world web-task tests, the score rose from 70.9 in the previous version to 75.0, while Astra scored 77.9。
The per-question model invocation costs listed in the report are:DeepSeek 0.22 USD，Astra 10.77 USD。
The report says that Anthropic recently tightened its safeguards, prompting domestic laboratories to turn their attention to Astra.
The report says that Liang Wenfeng was originally reluctant to rely on distillation in V4 research and development, but later changed his position due to competitive pressure within China.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This report links DeepSeek the improved capabilities of the new version to distillation using Astra outputs, and lists differences in web-task scores and inference costs; the claims come from unnamed sources and have not yet been officially confirmed or independently tested.

Why it matters to readers

If the report's content is accurate, it could affect external assessments of DeepSeek model capabilities, cost efficiency, and the boundaries protecting outputs from major AI models.

For beginners, the key point is: the report claims that the model's improved capabilities are related to distillation of Astra outputs, but this link has not yet been independently verified.

Risks and unknowns

DeepSeek Whether OpenAI and Moonshot AI will officially comment on the source of the training data is not stated in the evidence.
The evidence does not provide the names of the sources cited by The Information.
The evidence does not include the original training evidence or Browser Use the original testing methodology.
Browser Use It remains uncertain whether the test results can be reproduced by independent organizations.

Foreign media say DeepSeek and Kimi's new-model training made extensive use of GPT-6 Astra outputs

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center