Sometime between DeepSeek R1 in early 2025 and the steady drip of releases since, a quiet fact became impossible to ignore: the center of gravity for open-weights AI moved east. Not with one lab and one launch, but with a coordinated-looking surge across several companies, each with different strengths and the same appetite for shipping weights for free. If you're picking open models for a product and your shortlist is all American, you're ignoring where the best value in open AI actually lives right now.
The cast of labs
Start with the usual suspects. DeepSeek, the one that kicked the door down, continues to ship strong open reasoning and general models at prices that embarrass coastal API providers. Alibaba's Qwen team releases prolifically across sizes and ships most of it under Apache 2.0. Then there's Moonshot AI, maker of the Kimi series, which released Kimi K2 — a large mixture-of-experts model tuned hard for coding and agents. And Zhipu AI (Z.ai internationally), whose GLM family, including GLM-4.5 and its successors, keeps pushing capable open weights with strong tool-use behavior.
Add MiniMax, 01.AI and a rotating cast of smaller labs, and you get an ecosystem rather than a moment. These are real companies with real research budgets and real infrastructure, building on Chinese domestic accelerators partly because export controls on high-end NVIDIA hardware left them no other choice. Constraints bred brutal efficiency — when you can't brute-force a problem with compute, you get very good at architecture and data, which is exactly what you'd want a lab to optimize for.
Why they're winning on open weights
Three forces are converging. First, compute scarcity. When your lab can't casually order tens of thousands of H100s, you get savagely efficient at MoE routing, low-precision training and data curation. DeepSeek's published training economics made Western labs look profligate, and the shock forced everyone to rethink what enough compute really means.
Second, competition at home. Chinese labs are fighting each other for developer mindshare and enterprise adoption, and open releases are the fastest way to win both. Every strong open release is a recruitment ad, an API upsell and a national-prestige move at the same time. Third, policy. Open-sourcing models builds influence and talent pipelines, aligns with the push for technological self-reliance, and signals a mature AI industry rather than a copied one. The flywheel feeds itself: open weights attract the global developer community, whose feedback and fine-tunes feed back into the next release.
The result for the rest of us: frontier-adjacent open models with permissive or semi-permissive licenses, arriving faster than most Western teams can evaluate them. That's not a complaint — it's a queue management problem.
Kimi K2: the agentic one to watch
Kimi K2 landed as a very capable open MoE with on the order of a trillion total parameters and a far smaller active count, reportedly trained with a heavy emphasis on code and long-horizon tool use. Early independent results put it in genuinely useful territory for software engineering tasks — the kind of SWE-bench-adjacent performance that made engineers sit up, because those are the tasks that map to real product value.
It runs well in agent harnesses that need reliable function calling over long trajectories, and early adopters report it holding up through multi-step refactors and bug hunts in ways smaller models can't. It's not cheap to host — this is data-center-class hardware, full stop — but for teams building coding agents or autonomous workflows, Kimi K2 deserves a genuine evaluation rather than a skim of the announcement. If your agent routinely blows past forty tool calls before finishing a job, K2 is exactly the kind of model you should be benchmarking.
GLM: the pragmatic generalist
Zhipu's GLM line is less flashy and more useful than it has any right to be. The GLM-4.5 family mixes a reasoning mode and a fast mode, handles tool use competently, and ships open weights that the community quickly quantized and redistributed. If Kimi K2 is the big swing, GLM is the workhorse — the model you wire into a support pipeline or extraction service and forget about, which is the highest compliment a production model can receive.
GLM's tool-calling reliability is a particular strength; in agent loops, malformed JSON kills otherwise perfect runs, and this is one area where the GLM family has punched above its size for a while now. For teams self-hosting a single capable model for mixed general traffic, it's an underrated option that won't make headlines and absolutely will lower your bill.
Qwen: the default you'd already be using
We wrote a full breakdown on Qwen 3, so briefly: it's the most complete open-weights lineup out there, from laptop-sized dense models to large MoEs, mostly Apache 2.0, and quietly deployed everywhere from hobby projects to national infrastructure. If you only run one bake-off against this wave, Qwen is the baseline. Its smaller models dominate local-inference charts, and its larger ones compete at the top end — breadth is the moat.
DeepSeek: the spark that lit the fire
It's worth separating DeepSeek from the rest as its own story. R1's release in January 2025 proved that pure reinforcement learning could elicit frontier-grade reasoning from open weights, under an MIT license, at API prices nobody could match. V3 backed it up as a general model. Whatever you think of the politics, that one release broke the industry's assumptions about the cost of intelligence and made this whole wave inevitable — every lab on this list benefited from the price pressure and attention it created.
What to watch out for
Enthusiasm shouldn't be blind. A few honest caveats:
- Censorship and refusal behavior around Chinese political topics is baked into base weights across this ecosystem. Self-hosting and fine-tuning mitigate this; hosted APIs inherit it wholesale. Test adversarial prompts before you ship.
- Licensing varies by model and region. Apache 2.0 covers much of Qwen; other labs use custom licenses with commercial-use terms worth reading line by line, especially around redistribution and hosted services.
- Data and compliance questions. Open weights don't automatically answer questions about training-data provenance, data residency, or what happens to prompts sent through a hosted endpoint. Regulated industries should self-host or demand strong contractual answers, not vibes.
- Export-control politics. US policy toward these models can shift. Open weights can't be recalled once released, but hosted endpoints, enterprise support and GPU shipping logistics can all be disrupted.
- Quality variance across checkpoints. GLM and Kimi names cover several releases; older checkpoints differ meaningfully from current ones. Pin versions and re-evaluate.
Adoption patterns we're seeing
Three patterns repeat across the teams actually deploying this wave. First, self-hosted gatekeepers: a Chinese open model sitting inside the VPC, handling extraction and routing, with a US frontier API as escalation. Second, Chinese-market product teams who ship Qwen or GLM natively because their users and compliance reality live there, and because the models are genuinely better on Chinese-language nuance than Western alternatives. Third, cost-driven migrations at Western SaaS companies, moving bulk inference from premium APIs to self-hosted Qwen or DeepSeek and pocketing the margin.
Notice what's missing from that list: "we picked it because it's Chinese." Nobody serious is doing that. They picked it because the benchmark-per-dollar was best, and they'd have done the same with any lab's weights.
Running a fair bake-off across this wave
If you're going to evaluate Kimi K2, GLM and Qwen side by side — and you should — the details decide the outcome. Same prompts, same temperature, same quantization target, same hardware class, and crucially, the same subset of your real traffic rather than a leaderboard you downloaded. Score structured-output reliability and tool-call success alongside the "does the answer look smart" judgment, because the former predicts production behavior and the latter predicts Twitter engagement. Run at least a hundred examples per model; forty feels like enough until the third flip-flopped ranking.
Two procedural notes from painful experience. First, version everything: checkpoint hash, serving engine, CUDA stack, prompt template. These labs re-upload and rename checkpoints cheerfully, and "we tested GLM" means nothing three months later without the hashes. Second, blind your evaluators where feasible. Knowing which answer came from which lab leaks straight into scoring, especially on multilingual tasks where folk wisdom says one lab is "better at Chinese" — sometimes true, sometimes a story people tell themselves instead of measuring. A clean bake-off costs a week and buys you a defensible decision; a vibes-based one buys you a migration you'll redo next quarter.
The geopolitics, without the panic
Yes, export controls drove some of this efficiency. Yes, there are plausible scenarios where US policy restricts hosting or use of Chinese-origin models for certain government-adjacent workloads. But open weights are a diffusion technology — once released, they can't be un-released, and the genie is not going back in the bottle. The realistic near-term friction is procurement and compliance friction, not technical infeasibility. Plan for it the way you plan for any supplier concentration risk: dual-source, document provenance, and keep your eval harness portable.
The strategic read
Ten years ago, open-source infrastructure came from everywhere and nobody blinked. Now the most capable free models in the world mostly come from Chinese labs, and the closed frontier lives in San Francisco and London. That's not a doom narrative; it's a procurement reality you have to price in. If you're building on open weights and you ignore this wave, your open stack is a compute decision made in one country and a model decision made in another — whether you admit it or not.
The practical play is boring and effective: evaluate Kimi K2, GLM and Qwen 3 on your own workloads, self-host where the numbers agree, read the licenses, and keep US models in the mix too. The teams getting the most value from open weights right now aren't picking sides in a geopolitical framing. They're running the bake-off, watching the eval scores shuffle every quarter, and letting the benchmarks and the bill decide.
One closing observation: the health of this wave is good for everyone. Real competition at the open frontier forces every lab — American, Chinese, European — to ship better models for less. The winner isn't a country; it's whoever deploys first on their own data and learns hardest.