Trackers / Frontier model tracker

China Frontier LLM Release Tracker

Every flagship large language model from China's frontier labs: lab, release date, parameters, context, license, public pricing and a benchmark claim, each row traced to a named source.

As of 28 July 2026. Updated from China AI Dispatch's daily reporting. Machine-readable version:frontier-models.json.

The question behind every Chinese model release is the same: how close is it to the US frontier, and what does it cost to run? This page answers both as a table. It lists the flagship models from the labs that matter, with when each shipped, its size and context where the lab disclosed them, its license, its public price, and one load-bearing benchmark result, each attributed. It is a companion to the domestic chip scoreboard, which tracks the silicon these models increasingly run on.

Two honest caveats sit over the whole table. First, benchmarks are contested ground. Where a number is the lab's own claim, we mark it lab-reported; where it comes from an independent leaderboard (Artificial Analysis, OpenRouter, Chatbot Arena, Code Arena), we mark it independent. A vendor's "we beat everyone" is never presented as neutral fact. Second, several labs do not publish a parameter count or context figure, most notably Alibaba for Qwen; those cells read "n/d" rather than carrying a guess. We track one flagship per lab to keep the table readable; the daily carries every point release.

The tracker

Flagship large language models from Chinese frontier labs, newest first. Cells marked n/d are specs the lab did not disclose. Each benchmark cell is tagged with who reported it. Each row links to its source below.

Scroll the table sideways to see every column.

How to read this

Open weights are the through-line. DeepSeek V4, Zhipu's GLM-5.1 and GLM-5.2, Moonshot's Kimi K2.6 and K3, MiniMax M3 and Meituan's LongCat-2.0 all ship open weights, most under a permissive license. The lone closed flagship in the table is Alibaba's Qwen3.7-Max, which stays proprietary and API-only. When a lab cannot out-spend the US frontier on chips, giving the model away is how it sets the standard.

Price is the other weapon. DeepSeek V4-Pro lists at roughly one-seventh the price of Claude Opus 4.7; MiniMax sells about fifteen times the token volume of a comparable Claude plan for the same money; GLM-5.2 runs at roughly one-sixth the API cost of GPT-5.5. The models are close on capability and far cheaper to run, which is the entire distribution thesis.

The benchmarks are close but not settled. On independent leaderboards the best Chinese models now land between GPT-5.5 and the top Opus builds: on the Artificial Analysis index GLM-5.2 was, by Artificial Analysis's own reading, the first open-weight model to reach 51, and Kimi K3 topped the Frontend Code Arena. But the same independent reporting is clear that Chinese models still trail the top US labs on the hardest coding benchmarks, and the "we beat everyone" figures are the labs' own. For the chips these models run on, see the domestic chip scoreboard.

Who is actually using these models

Increasingly, American enterprises. An IDC survey published in late July 2026 found that 47% of decision-makers at US firms with more than a thousand employees had put a Chinese model into at least one use case, and about one in five said they use one heavily. At a point in mid-July, all five of the top models on OpenRouter, the marketplace that routes enterprise traffic to whichever model wins on price and performance, were Chinese. (IDC and OpenRouter, via The Next Web and China AI Dispatch, 28 Jul 2026.)

The named switchers are not lab experiments. Coinbase moved work onto Moonshot's Kimi and Zhipu's GLM and says it halved its AI spending; DoorDash routes lower-level work to Kimi; Airbnb runs customer service on Alibaba's Qwen; and Cursor, the coding tool, built its Composer 2 model on Kimi foundations. The pattern is consistent: keep an expensive US model for the hardest slice of work and hand the rest to a Chinese model that costs a fraction as much. (China AI Dispatch, "The Trade-Down," 28 Jul 2026.)

The cost gap is structural, not a promotion. UBS estimates the leading Chinese models cost roughly a tenth as much to train as comparable US systems, with API prices at 10 to 20% of the foreign alternative, reached through smaller and more sparsely activated mixture-of-experts designs, GPU utilization above 70% against an industry norm near 40%, and cheaper power. On one legal benchmark the spread was stark: about $2.40 per task for GLM-5.2 against about $31 for Anthropic's Fable, roughly thirteen to one. (UBS via TechNode; Applied Compute's Harvey case study; via China AI Dispatch.)

The caveat cuts both ways. Chinese commentators are not triumphant about it: Caixin argues the cheap-AI story runs into the heavy-industrial cost of actually serving trillion-parameter models at scale, and selling intelligence at cost is not a business. And both Washington and Beijing are now weighing export controls on advanced models and weights, so the open window in which a US company can download a Chinese model and run it in production may not stay open indefinitely. (Caixin and the Financial Times, via China AI Dispatch, 28 Jul 2026.)

Common questions

Sources and method

Every row traces to a named source carried in China AI Dispatch's reporting: DeepSeek, Zhipu, Moonshot, MiniMax and Alibaba's own technical reports and launch posts for specs and lab-reported benchmarks; Reuters, SCMP, InfoQ, tmtpost, 36Kr, Qbit and Geek Park for release facts; and Artificial Analysis, OpenRouter, Chatbot Arena, LMArena and Code Arena for independent leaderboard placements. Where a row links to a CAD issue, that issue cites the underlying primary source.

Two classes of figure are handled with care. Benchmark results are tagged by who produced them: a lab-reported tag means the number is the vendor's own claim and should be read as such; an independent tag means a third-party leaderboard. Parameter counts and context windows are only listed where the lab disclosed them; Alibaba does not publish Qwen's, so those cells read "n/d" rather than carrying an outside estimate. We track one flagship model per lab to keep the table legible, so a lab's earlier or lighter variants are not all listed here.

This is a living document, updated from the daily newsletter, which always carries the newest releases first.