China's AI Labs Turn Open Weights Into Four Different Strategies

Column Overview

On July 25, Nvidia CEO Jensen Huang used his first-ever post on X to put his name behind something unusual: a joint letter, co-signed with Meta, Microsoft, and Hugging Face among others, arguing that open-weight models deserve to sit alongside closed frontier systems as a permanent part of the AI landscape rather than a lesser, catch-up tier. It read as a statement of principle more than an announcement of any product. But the timing gave it an odd resonance, because in the days on either side of that letter, four Chinese AI labs put out four releases that, taken together, make the letter's abstract point concrete — and complicate it. Moonshot AI, DeepSeek, Alibaba, and MiniMax are all shipping open weights right now, and they are not doing it for the same reason or in the same way. "Open weight" is not a single strategy these companies have converged on; it is a label that four different companies are filling with four different bets.

A Letter That Says the Quiet Part Out Loud

What made Huang's letter notable wasn't the argument itself — plenty of researchers have made the case that open weights accelerate downstream research and give smaller developers a fair shot at building on frontier-grade technology. It was who signed it and when. A hardware company, a foundation-model lab, a cloud platform, and an open-model hosting hub agreeing publicly that open weights are not a stopgap but a durable category is a signal that the commercial ecosystem around open models — chips, hosting, fine-tuning services, enterprise deployment — has matured enough to be worth defending in writing. That ecosystem is exactly where Chinese labs have been the most active over the preceding months, which is what makes their near-simultaneous releases worth reading side by side rather than as isolated news items.

Moonshot's Bet: Bigger Is the Point

Moonshot AI has been the most literal about what "open weight" can mean as a marketing and technical claim. It released a preview version of Kimi K3 on July 19 at 2.8 trillion parameters with million-token-scale memory, then followed on July 28 with a full release billed as the largest open-weight model released to date, positioned as rivaling top closed U.S. systems on capability. Whether or not "largest" survives scrutiny as benchmarks are independently reproduced, the strategic logic is legible on its own terms: Moonshot is treating raw scale as a form of open-weight credibility. If you are going to give a model away, the argument goes, give away one big enough that nobody can dismiss the open tier as a discount version of the real thing. It's a bet that scale itself is the headline, and that headline is worth more to Moonshot right now than a leaner, more efficient model would be.

DeepSeek's Two-Track Play: Give Away the Weights, Undercut on Price

DeepSeek is running a different experiment, and it's worth noticing that it has two moving parts rather than one. In April, the company released an open preview of DeepSeek V4, a trillion-parameter model with million-token context — comfortably in the same scale conversation as Moonshot's release. But a month later, DeepSeek did something that has nothing to do with open weights at all: it made a 75% discount on its flagship V4-Pro API pricing permanent, rather than a limited-time promotion, effectively making it the cheapest frontier-grade API on the market. Put those two moves together and DeepSeek's strategy comes into focus as two-pronged rather than singular. Open weights serve the audience that wants to self-host, fine-tune, or audit the model directly; aggressive API pricing serves the much larger audience that just wants inference without any of that overhead. DeepSeek isn't choosing between "open" and "cheap" as competing identities — it's using each to reach a different customer who wouldn't have been served by the other.

Alibaba's Portfolio Approach: One Flagship, One Specialist

Alibaba's Qwen team has taken a different route again, treating open weights less as a single flagship bet and more as a portfolio to be filled out deliberately. Qwen3-Coder, a 480-billion-parameter model released specifically for programming tasks, and Qwen3-Max-Preview, a trillion-parameter general-purpose flagship, arrived within about six weeks of each other and were clearly not meant to compete with one another — they're meant to cover different jobs. That's a meaningfully different open-weight strategy than "release the biggest thing you have." It reads more like a product line than a single product: a generalist model for the broadest possible developer base, and a specialist model for the one vertical — code generation and agentic coding workflows — where enterprise willingness to pay for marginal quality gains is highest. Coverage, not scale, is the organizing idea.

MiniMax's Narrower Wager: Agents That Improve Themselves

MiniMax has picked the narrowest lane of the four, and arguably the most technically specific one. M2.7, released as open weight in April, is pitched around a self-evolving agent architecture rather than general chat or reasoning ability, with reported scores of 56.22 on SWE-bench Pro and 57.0 on Terminal-Bench 2 — benchmarks that measure an agent's ability to complete real coding and terminal-based tasks autonomously, not just answer questions well. Where Moonshot is selling scale and Alibaba is selling coverage, MiniMax is selling a capability: an open model built around the idea that an agent should get better at a task through its own iteration rather than needing a bigger base model or more fine-tuning data. It's a bet that the next competitive axis in open-weight models won't be parameter count at all, but how well a model functions as the core of an autonomous system.

What "Open Weight" Actually Buys Each of Them

Line these four up and the interesting conclusion isn't that Chinese AI labs have all decided open weights are the future — it's that they've each concluded something different about what open weights are for. For Moonshot, open weights are a scale claim that doubles as a credibility claim. For DeepSeek, they're one half of a two-track distribution strategy where price does the other half of the work. For Alibaba, they're a way to build market coverage across a general flagship and a coding specialist without having either compete against itself. For MiniMax, they're a vehicle for a much narrower and more technical bet on agentic self-improvement. Jensen Huang's letter treated "open weight" as a category worth defending in principle. What these four releases show is that inside that category, the actual strategic content varies enormously — and that variety, more than any single release, is probably the more durable story about where open models are headed next.