OpenAI-backed Harvey — the legal AI unicorn valued at $11 billion, serving 1,300+ institutions and 100,000 lawyers — just shipped its first model built on an open-weight base. The base is not from OpenAI, Anthropic, or Google. It's Kimi K3, the latest open-weight model from China's Moonshot AI.
Harvey is not the exception. It's the pattern. Over the past six months, an unusually long list of American AI products quietly switched to Chinese open-weight models as their foundation: Cursor's Composer 2 and 2.5 run on Kimi K2.5; Devin's SWE-1.7 is built on Kimi K2.7; London's Cosine fine-tuned Kimi K2.6 into Lumen Outpost; and Mira Murati's Thinking Machines bootstrapped Inkling with DeepSeek V3's architecture and Kimi-generated synthetic data. Every new Kimi generation now seems to spawn another US-branded model.
"Made in America" is becoming "prefabricated in China, assembled in America." That isn't a punchline — it's a supply-chain structure. Here's what is actually happening, why it matters, and what it changes about how you should build.
Why an $11B legal unicorn skipped GPT for Kimi
Start with the problem Harvey is solving. Courts are getting burned by general-purpose AI: a judge's clerk used Perplexity to draft a ruling, and the filing came back with fabricated citations and the wrong party names; another litigant hid white text inside a court document instructing any AI reading the file to rule in his favor. Legal work exposes what generic models can't do — the answer isn't the deliverable, the workflow is.
A lawyer doesn't receive a multiple-choice question. She receives a partner's request and a pile of client files with truth and traps mixed together. The model has to decide where to start, search and read continuously, map risk to citation, then produce a clause-level redline or a memo that can actually be delivered. That judgment — when to pull one more document, why a deliverable is still wrong — lives in thousands of real executions, not in a prompt.
That's why Harvey built Tenet, its first open-weight post-trained model, aimed at long-horizon, tool-calling legal tasks. On Harvey's own legal-agent benchmark (LAB), Tenet completed nearly twice as many tasks as its base model. The training data, the workflow scaffolding, the evaluation — that's Harvey's IP. The raw reasoning substrate is Kimi K3.
The relay: each Kimi generation becomes a US product
Harvey is the latest, not the first. Read the list as a handoff:
- Cursor → Composer 2 / 2.5 (Kimi K2.5): the hottest AI coding tool in the Valley, valued in the tens of billions, built its own model on Kimi's open weights after being pressed about its origin — and Composer 2.5 kept the same base.
- Devin → SWE-1.7 (Kimi K2.7): Cognition's famous "first AI software engineer" now openly uses Kimi K2.7 as the foundation for autonomous coding work.
- Cosine → Lumen Outpost (Kimi K2.6): a London startup that maintains decades-old legacy code for banks and airlines fine-tuned Kimi K2.6 to cut hallucinated "garbage code" and solve problems proactively — a small company that can't outspend OpenAI, so it bought a base model instead.
- Thinking Machines → Inkling (DeepSeek V3 architecture + Kimi K2.5 synthetic data): Murati's $12B+ startup referenced DeepSeek V3's architecture for its foundation model and used Kimi-K2.5-generated synthetic data to kick off supervised fine-tuning.
None of these companies are "just rebranding." They invest real data, compute, and domain experience. But they've stopped pretending a model was trained from scratch — because bloodline doesn't matter when cost does.
From two paths to three: open weights become the upstream
For three years, AI startups had two options. Train a foundation model from zero — a billion-dollar bet only a handful can afford. Or rent a closed API (GPT, Claude) per token, which means the more your product succeeds, the more of your margin flows upstream, and you never own the model.
Chinese open-weight releases opened a third path: take a trained base, add your data and your process knowledge, post-train it for your business — and own the result. The open-weight model is the raw material; the startup is the factory that turns it into a product.
The right analogy isn't "open-source software" — it's semiconductors. Design in the US, fabrication in Taiwan, assembly everywhere: a layered industry where each stage sells to the next. Chinese open-weight models are becoming AI's upstream industrial goods — the foundry layer. You can dislike the geopolitics, but the economics is structurally familiar: the commodity layer earns thin margins, the application layer earns the brand, and the customer often never knows which layer made the silicon. Most users of these products don't know the base model is Chinese. That's precisely the point — it's infrastructure now.
What this means for the market
Three consequences follow.
1. The moat moved. For application companies, the defensible layer is no longer "we trained our own model" — it's data, workflow, evaluation, and distribution. Harvey's LAB benchmark is worth more than Tenet's weights. That's a structural shift in where value accrues.
2. Closed labs are getting squeezed at the margin. Every task an open-weight fine-tune handles cheaper is a task removed from GPT/Claude's per-token meter. The frontier keeps them relevant, but the middle of the market is being commoditized — OpenAI's own portfolio company just picked a Chinese base over its own API.
3. Chinese models export influence through usage, not sales. Instead of selling API access, open-weight releases become de facto standards adopted inside Western products. Usage is the distribution channel; every downstream product that ships is a deployment of that ecosystem.
What to do about it
- Benchmark your task distribution, not the leaderboard. Harvey didn't pick Kimi because of a ranking — it built LAB to measure task completion inside its own workflow and found the post-trained model finished nearly 2x more. Run that experiment on your own data before choosing a base.
- Treat open weights as the default for narrow, high-frequency jobs. For bounded tasks — legal document review, legacy code, regulated verticals — domain fine-tunes on open bases are beating rented frontier APIs on cost and control.
- Reserve frontier APIs for open-ended work. The commodity layer handles the repeatable; the frontier handles the novel. Hybrid is the architecture.
- Read the license, then build the eval. Open-weight isn't free in every sense — check commercial terms, then build your eval harness first so you can swap bases when a better one lands.
The takeaway: the question is no longer "whose model is best?" — the commodity layer settles that on price. The question is "what can you post-train that they can't?"
