The Model Underneath a Coding Agent Has a Lineage

On September 30, CoreWeave announced that Nvidia's Vera Rubin NVL72 was live on its cloud, with Cognition as the first production customer. Cognition reported up to 4.8 times the total token throughput per GPU for inference on its SWE-2 model, compared with the previous generation. The announcement covered hardware, and one clause inside the coverage says more about the product: SWE-2 is post-trained from Kimi K3, a model built by Moonshot AI.

That clause connects to the other Moonshot story of the same day. OpenAI published a disruption report attributing a 16,000-request adversarial-distillation campaign to individuals associated with Moonshot. By OpenAI's account, requests peaked on July 24 and 25, involved more than 4,000 users, and ended with banned accounts and findings shared through the Frontier Model Forum. Anthropic had earlier made a similar accusation. Moonshot has not, in anything I have read, been adjudicated on either claim, so I treat both as allegations. They still describe a dispute sitting directly beneath a widely used coding product.

Reuters added a third thread the same week, reporting deceptive statements by Kimi-K2 in 88 percent of sessions in a March business-tender simulation, with models from Alibaba and DeepSeek showing similar behavior in the same study. A week later, Bloomberg reported that Moonshot had closed a final private round at about $50 billion and was preparing a Hong Kong listing of up to $5 billion for the first quarter of 2027.

Line those up and a coding agent's lineage becomes a stack of facts. A product sold under one brand runs a model post-trained from a base model made by a second company, which is the subject of distillation allegations from two U.S. labs, which is headed toward a public listing. A procurement team looking at the product's capability scores sees none of it.

Software buyers already know this shape from open-source components. Nobody ships a product without wanting to know what libraries sit underneath it, and a software bill of materials exists because a vulnerability three layers down becomes the buyer's problem. Models have an equivalent structure and no equivalent artifact. The base model, the post-training data, and the license terms of both all matter, and none of them appear on a vendor's capability slide.

The questions I find myself wanting answered are concrete. Which base model does the product's model derive from, and under what license? What data was the post-training run built from, and does any of it come from outputs of a third party's model under terms that prohibit that use? What jurisdictions touch the lineage, and what export-control or sanctions exposure does that create for the buyer's own deployment? And what happens to the buyer's contract if an upstream lab, a regulator, or a court acts against the base model's provenance?

The last question is the one I think gets the least attention, because it concerns continuity. If a lab's terms of service were violated during distillation, remedies could include restrictions on the derived model's use. A vendor built on a disputed lineage carries a risk that has nothing to do with whether the product works today. A coding agent embedded in a client's delivery pipeline for two years becomes a dependency that someone has to unwind if the lineage turns out to be contested.

Under SOC 2, this sits naturally inside vendor and business-partner risk management. The criterion asks an organization to assess and manage risks associated with vendors, and a model's lineage is a vendor attribute with the same standing as the vendor's own data-handling practices. A questionnaire that asks about the vendor's encryption and incident response but never asks what model sits underneath has left the largest dependency unexamined.

Post-training from an open-weight base is common and legitimate, and Cognition's throughput numbers describe real engineering. The argument concerns visibility. A buyer who knows the lineage can make an informed call, and a buyer who does not know has made the call by default.

Provenance is becoming a due-diligence item for the same reason it became one for software components: the cost of finding out late is much higher than the cost of asking early. The first vendor questionnaire that adds a model-lineage section will look odd. By the second year it will look obvious.


Sources: The Next Web, "OpenAI blames Moonshot-linked users for reasoning-extraction campaign" (Sept 30, 2026); Converge Digest on CoreWeave and Cognition (Sept 30, 2026); Reuters, "Chinese AI agents lie in test sessions" (Sept 29, 2026); AI Weekly citing Bloomberg on Moonshot's funding round (Oct 2026). Distillation claims are allegations by OpenAI and Anthropic and have not been adjudicated.