Two of China's top AI labs shipped upgraded open-weight models this month, and both reportedly leaned on outputs from OpenAI's GPT-6 Astra
Two of China’s most closely watched AI labs released new models in October. Both reportedly had help from an unlikely tutor: OpenAI.
Moonshot AI and DeepSeek are said to have used outputs from OpenAI’s GPT-6 Astra to develop the models they shipped this month. That puts a familiar question back on the table. When a lab learns from a rival’s answers, is that clever engineering or something closer to copying homework?
What the two labs actually released
DeepSeek went first. It launched DeepSeek-V4.1-Flash-1017 on October 17, 2026.
The update focuses on sharper task performance and better agentic coding. Agentic coding means the model doesn’t just suggest code. It plans and carries out multi-step programming work with less hand-holding.
DeepSeek kept its signature approach intact. The model stays open-weight and low-cost, which is the formula that made the lab famous in the first place.
Moonshot followed on October 23, 2026, with Kimi K3.1 Pro. It’s a substantial upgrade over a predecessor that came out only a few months earlier.
The model packs approximately 2.8 trillion parameters. Parameters are the internal dials a model tunes during training, so more of them generally means more capacity to learn patterns.
On the Artificial Analysis Intelligence Index, Kimi K3.1 Pro scored 55. That placed it fourth overall in the rankings.
It also beat GPT-6 Astra on certain agentic benchmarks, including Terminal-Bench 4.0. So a model that reportedly learned from Astra’s outputs is now outscoring Astra on some tests.
Independent evaluations put Kimi K3.1 Pro at or near the top of open-weight rankings. That makes it a credible alternative to closed US systems, at least on the benchmarks that measure this kind of work.
Both labs are pitching their models at practical jobs. Financial analysis and coding are the headline use cases.
The distillation question, again
The technique at issue is usually called distillation. Think of it like a junior chef tasting a master’s dishes over and over until they can reproduce the flavors without ever seeing the recipe.
Technically, it means feeding a strong model lots of prompts, collecting its responses, and using those responses to train a different model. The result can absorb much of the teacher’s capability at a fraction of the training cost.
This is not the first time these two names have come up in that conversation. In February 2026, Anthropic accused both Moonshot and DeepSeek of running distillation attacks on its Claude models.
Anthropic’s claim went further than borrowing answers. It alleged the labs used fraudulent accounts to extract outputs from Claude.
Why the money is watching closely
These aren’t scrappy side projects. Moonshot AI’s valuation reportedly reached around $50 billion by October 2026, ahead of a planned IPO in Hong Kong.
DeepSeek has been valued even higher in recent funding rounds. Between the two, that’s a lot of capital riding on the idea that Chinese labs can compete at the frontier.
The performance numbers support that pitch. A fourth-place finish on a major intelligence index, plus wins over a flagship OpenAI model on specific tests, is the kind of data point that shows up in IPO decks.
The distillation reports complicate it. If a lab’s edge depends partly on outputs from a competitor’s model, investors have to ask how durable that edge really is.
What this means
For Moonshot specifically, the stakes are immediate. A Hong Kong IPO at a valuation around $50 billion invites detailed questions about training practices, and distillation allegations are the type of thing prospective shareholders tend to notice.
For DeepSeek, the low-cost, open-weight playbook has been its calling card. If questions about where its training signal comes from start to stick, that cost advantage may draw more scrutiny than praise.
The thing to watch is OpenAI’s response. Anthropic chose to go public with its accusations. If OpenAI does the same over Astra, the debate over who owns a model’s outputs could move from industry gossip to formal policy.
