In August 2026, a legal technology company backed by OpenAI made an announcement that should have been embarrassing for OpenAI. Harvey, a legal AI platform used by thousands of law firms, revealed that its new flagship model wasn’t built on OpenAI’s technology, or Google’s, or Anthropic’s. It was built on top of a free, publicly downloadable model released by a Chinese startup most Americans had never heard of. Harvey, valued at $11 billion, built its new proprietary model on the open-source Kimi K3 model from Chinese company Moonshot AI, rather than the GPT technology from its own investor OpenAI.
This wasn’t a one-off. It’s the leading edge of a much bigger shift that has real implications for anyone trying to make money with AI in 2026 — not just billion-dollar startups, but freelancers, solo app builders, and small agencies from Bangkok to Jakarta to Ho Chi Minh City.
What “Open-Weight” Actually Means (And Why It’s Different From “Free”)
To understand what’s happening, you need one piece of jargon: open-weight. A closed AI model, like the ones sold by OpenAI or Google, only lets you send it a question through an app or an API (a paid connection that lets your own software talk to someone else’s AI) and get an answer back. You never see how it works inside. An open-weight model is different: the company publishes the actual trained “brain” of the model — the weights, meaning the millions of internal settings the AI learned during training — so that anyone can download it, run it on their own servers, and modify it. It isn’t fully “open-source” in the traditional software sense, because the raw training data and code usually stay private. But it’s open enough that a small team, or even one person, can build a specialized product on top of it without paying a big AI company for every single request.
China’s AI labs have leaned hard into this model. Moonshot AI, a Beijing-based lab, releases a family of open-weight models called Kimi. DeepSeek, another Chinese lab that first grabbed global attention in early 2025, does the same with its own model family. Both have become go-to foundations for companies that want frontier-level AI performance without frontier-level bills.
The Legal Startup That Bet Its Future on a Chinese Model
Harvey’s case is the clearest example of what’s now called post-training: taking someone else’s open-weight base model and continuing to train it on your own specialized data, so it gets much better at one narrow job. Harvey completed its post-training in about two months using roughly 150 Nvidia B300 GPUs, lifting its pass rate on its in-house legal benchmark from around 11 percent to nearly 20 percent, while cutting inference costs to under a quarter of closed-source frontier models. The resulting model, called Tenet, completes almost twice as many held-out legal tasks as the base Kimi K3 model, while contract drafting, review, and negotiation work saw completion rates rise by around 20 percent. Kimi K3 itself is a massive model, with 2.8 trillion total parameters and support for context windows up to one million tokens (a “token” is roughly a word-piece of text, so a million tokens is enough to feed in an entire book or codebase at once).
The commercial logic is blunt: every time Harvey used a rival’s AI model, that rival collected the invoice. Owning a fine-tuned version of an open-weight model turns that ongoing expense into something Harvey controls itself.
A $29 Billion Coding Startup Got Caught Doing the Same Thing
Cursor, a wildly popular AI coding assistant, tried a quieter version of the same move — and got caught. When Cursor launched its “Composer 2” coding model in March 2026, developers noticed a stray internal model name in the company’s own traffic that gave the game away. Composer 2 wasn’t a proprietary breakthrough at all; it was built on top of Kimi K2.5, an open-weight model from Moonshot AI, and Cursor confirmed the connection only after it had already leaked. The backlash wasn’t about whether building on an open-weight model was legitimate — it clearly is — but about Cursor not disclosing it upfront.
Two months later, Cursor doubled down anyway. Composer 2.5 shipped on the same Kimi K2.5 checkpoint as its predecessor, but trained on 25 times more synthetic training tasks, with the vast majority of the compute budget spent on Cursor’s own reinforcement-learning pipeline. The payoff is price: the open-weight K2.5 base is the reason Cursor can price Composer 2.5 at roughly one-tenth the rate of a comparable closed frontier model, while still delivering competitive coding performance.
Even a Former OpenAI Chief Technology Officer Is Building on Chinese Architecture
If you assumed this was purely a scrappy-startup habit, consider Thinking Machines Lab, the AI company founded by Mira Murati after she left her role running technology at OpenAI. In July 2026 the company released its first big model, called Inkling. Inkling uses a mixture-of-experts design — a common architecture that only activates a fraction of its total parameters for any given task, keeping very large models faster and cheaper to run — with 975 billion total parameters but only about 41 billion active at once. Thinking Machines Lab was explicit about where the design came from: Inkling’s architecture references China’s DeepSeek-V3, and part of its post-training used synthetic data generated by Moonshot AI’s Kimi K2.5.
That’s three separate American companies — a legal-tech unicorn, a coding-tool giant on the edge of a reported acquisition talks, and a lab run by one of the most credentialed executives in AI — all choosing to stand on a Chinese foundation rather than build entirely from scratch or lease a closed competitor’s model.
Why This Isn’t Just a Big-Company Story
Here’s the part that actually matters if you’re trying to earn a living with AI rather than run a startup: the economics that make open-weight models attractive to a $15 billion legal-tech company scale down perfectly to a one-person operation. The pricing gap is enormous. DeepSeek’s current API centers on two models, with the everyday option priced at roughly $0.14 per million input tokens and $0.28 per million output tokens, both with a one-million-token context window. Compare that to premium closed models from Western labs, which frequently charge many times more per million tokens for comparable work.
แนะนำ
Independent developers are already organizing around this gap. One solo founder building tools for other bootstrapped developers described 2026 as an “AI budget crunch,” where solo founders can’t afford to lock themselves into one expensive AI provider, nor can they afford to ignore the price-to-performance advantage of Chinese models like DeepSeek and Qwen. That founder built a routing service specifically so other indie developers could plug into cheap Chinese open-weight models without dealing with region-specific billing headaches.
This is, in miniature, exactly the Harvey playbook: pick a narrow, well-defined job (contracts, customer support replies, product descriptions, tutoring in a specific subject), take a cheap open-weight base model, and fine-tune or carefully prompt-engineer it to be excellent at that one thing. You don’t need $2 billion in venture funding to do this. You need a laptop, a few hundred dollars of API credit, and a genuinely narrow niche where “good enough, ten times cheaper” beats “state-of-the-art but expensive.”
What This Looks Like on the Ground in Thailand and the Rest of Southeast Asia
This trend is arriving in a region that’s unusually well-positioned to benefit from it. Indonesia and Vietnam together already account for more than 10 percent of global freelance-platform registrations, and 78 percent of freelancers globally now use AI tools to speed up their work. Cheap, open-weight-powered AI tools lower the entry cost for exactly that kind of freelance and micro-business work.
For readers actually based in Thailand, a few local details are worth knowing. LINE OA (LINE Official Account) is the business-messaging tool most Thai companies use to chat with customers, similar to a WhatsApp Business account. PromptPay is Thailand’s real-time bank-to-bank payment system, used the way Americans might use Venmo or Zelle. Fastwork is a Thai freelance marketplace, roughly Thailand’s answer to Upwork, where local freelancers sell design, writing, and increasingly AI-assisted services. And Shopee and Lazada are the two dominant e-commerce marketplaces across Southeast Asia — many small sellers already use AI to write product listings and translate them across the region’s languages. An AI-powered listing-writing or customer-chat tool built cheaply on an open-weight model and sold through Fastwork or bundled into a LINE OA setup is a very direct, very local version of the Harvey-to-Cursor pattern playing out at enterprise scale.
Thailand has also made it easier for foreign freelancers to actually live there while doing this kind of work. The Destination Thailand Visa, introduced for freelancers, entrepreneurs, and remote workers, grants multiple entries over five years, with individual stays of up to 180 days extendable to 360 days, and requires applicants to show a bank balance of roughly 500,000 Thai baht, about $15,000. Combine that kind of long-stay flexibility with sub-cent-per-request AI costs, and the barrier to launching a viable one-person AI service from a co-working space in Chiang Mai or Udon Thani keeps shrinking.
The Catches Nobody’s Advertising
None of this is a free lunch. Cursor’s undisclosed use of Kimi K2.5 shows that open-weight licenses often carry attribution requirements that companies can violate, intentionally or not — a risk for anyone reselling a fine-tuned model commercially without checking the license terms. Benchmark numbers from any company, including Harvey’s own reported gains, are self-reported and can be optimized for the test rather than real-world performance, so treat vendor claims with some skepticism. There’s also a geopolitical dimension: several governments and regulated industries restrict the use of Chinese-origin models for compliance or security reasons, so a tool built for regulated sectors like law, healthcare, or finance needs extra diligence.
The Takeaway
The most valuable AI skill in 2026 isn’t training your own frontier model from scratch — that’s a billion-dollar game reserved for a handful of labs. It’s the skill Harvey, Cursor, and Thinking Machines Lab all demonstrated at massive scale: taking a capable, cheap, open-weight foundation and narrowing it, ruthlessly, into something excellent at one specific job. Whether you’re a freelancer in Udon Thani, a digital nomad on a Destination Thailand Visa, or a solo developer anywhere in the world, that same move — cheap base model plus narrow specialization plus a real customer problem — is now available to you for the cost of an API key, not a venture round.
Key Takeaways
- Three major AI companies — Harvey, Cursor, and Mira Murati’s Thinking Machines Lab — built their flagship 2026 products on top of free, Chinese-made open-weight AI models rather than starting from scratch.
- Harvey’s legal AI model, Tenet, nearly doubled its task-completion rate after two months of specialized training on top of Moonshot AI’s Kimi K3.
- Cursor was caught quietly building its Composer coding models on Kimi K2.5, later using the same base to undercut rival pricing by roughly 90 percent.
- Open-weight models from labs like DeepSeek now cost a small fraction of premium closed AI models, putting the same strategy within reach of solo freelancers and small businesses.
- Southeast Asia, including Thailand’s freelance and digital-nomad-friendly visa policies, is positioned to benefit directly from this shift toward cheap, specialized AI tools.
Frequently Asked Questions
แนะนำ
Q: Can you actually make money with AI tools in Thailand in 2026?
A: Yes — the same strategy powering billion-dollar startups (taking a cheap open-weight AI model and specializing it for one job) works at freelance scale, and Thailand’s freelance platforms and digital nomad visa policies make it practical to run this kind of business locally.
Q: What is an open-weight AI model, in simple terms?
A: It’s an AI model whose trained internal settings are published publicly, so anyone can download it, run it on their own computers or servers, and customize it, unlike closed models that you can only access through a paid API.
Q: Is Kimi the same thing as DeepSeek?
A: No, they’re separate open-weight model families from two different Chinese AI labs — Kimi comes from Moonshot AI and DeepSeek comes from the company of the same name — though both compete on similar price-to-performance ground.
Q: Why would a company backed by OpenAI use a Chinese AI model instead?
A: Cost and control were the main drivers for Harvey; building its own model on an open-weight base cut inference costs sharply and reduced its dependence on paying its own AI-provider investors for every request.
Q: Is it legal for a US company to build products on Chinese open-weight models?
A: Generally yes for open-weight models released under permissive licenses, though some licenses require attribution above certain usage thresholds, and separate rules restrict Chinese-origin AI models in specific regulated industries or government contexts.
Q: How much cheaper are open-weight models like DeepSeek compared to ChatGPT or Claude?
A: Current open-weight options can run several times cheaper per unit of text processed than premium closed models from Western labs, though exact pricing shifts frequently as providers adjust rates.
Q: What is “post-training” and why does it matter for making money with AI?
A: Post-training means taking an already-trained open-weight model and training it further on your own narrow, specialized data, which is exactly how a solo builder can turn a general-purpose free model into a paid, specialized tool.
Q: Do I need to know how to code to build a business this way?
A: Not necessarily for the simplest versions — many freelancers get value from skillfully prompting an existing open-weight-powered tool for a specific market, though building and reselling a genuinely fine-tuned model does require technical skill or a technical partner.
Q: What is the Destination Thailand Visa and does it help with this kind of work?
A: It’s a long-stay visa aimed at freelancers, entrepreneurs, and remote workers that allows multiple entries over five years and stays of up to 360 days, making it easier for foreign AI-tool builders to base themselves in Thailand while running an online business.
Q: Are Vietnam and Indonesia also good markets for AI-based freelance work?
A: Yes — both countries already account for a large share of global freelance-platform sign-ups, and a majority of freelancers in the region now report using AI tools to work faster and take on more projects.
Q: What are the biggest risks of building a business around open-weight AI models?
A: Vendor benchmark claims can be inflated, some open-weight licenses carry attribution obligations that are easy to overlook, pricing on “cheap” models can change with little notice, and regulated industries may restrict which models you’re allowed to use at all.
Q: What’s a realistic first step for someone who wants to try this?
A: Pick one narrow, repetitive task you or people you know deal with regularly, test whether a low-cost open-weight model can handle it well through a paid API, and only invest in deeper fine-tuning once you’ve confirmed real demand for the specialized version.
