We run the models, publish the logs, and explain what the scores actually mean.
Moonshot AI opened Kimi K3 full weights on July 27, 10:00 GMT. Within 72 hours: 180K Hugging Face downloads, more than 340 derived fine-tunes, and official adapters from vLLM, SGLang and Llama.cpp. …
Alibaba's Qwen 3.8 was wired into OpenRouter, OpenCode, Hermes Agent and Vercel within hours of release — the same Western toolchain adoption pattern DeepSeek saw with V4 Flash.
MiniMax's H3 video model reached #1 on Hugging Face trending, passing DeepSeek V4 Flash in community attention — a first for a Chinese video-generation model in the open-weights space.
Per Bloomberg, the White House signaled in a closed-door meeting that Chinese open-weight AI models would be exempt from US security review requirements — a policy stance directly affecting self-hosted DeepSeek/Qwen/Kimi deployments.
In a CNBC interview, Hugging Face's CEO said China is beating the West on open-weight models — the latest mainstream acknowledgment of a gap that has been building since 2025.
2 formally tested in-house with raw logs, 5 smoke-tested (N=1), 3 compiled rows with every number sourced, 4 international reference — 6 tracked without scores. Last updated: 2026-08-08.
We compiled the same 11-task battery on 9 frontier models: GLM-5.2, Seed 2.1, Step 3.5 Flash, Hunyuan A13B, Ling 2.0 vs GPT-5.6 Luna Pro, Claude Opus 5, Gemini 3.6 Flash, Grok 4.5. Six scored 11/11; the only US-side failure was a quota issue, not capability. Costs spanned ~100x, from $0.0004 to $0.04. Raw logs public.
MiniMax H3's weights are on Hugging Face, and the license excludes the EU, UK, South Korea, and the US. The license text, verbatim, and what it means for developers.
At mid-2026 the China-US model race has a scorecard with two columns: China wins open weights, volume, and price; the US still holds the frontier lead. 13 straight weeks of Chinese token volume, an 80% OpenAI price cut, and the numbers in between.
Zhipu's GLM-5.2 leads the open-source field on Artificial Analysis at 51 points, tops the Code Arena for available models, and publishes its weak spots: 45-minute reasoning runs and a 5% knowledge lag. We compiled the numbers.
Moonshot's Kimi K3 is the largest open-weight model ever shipped: 2.8T params, 180K Hugging Face downloads in 72 hours, 340+ fine-tunes, and a demand crush that paused new subscriptions. We compiled the numbers and the caveats.
MiniMax H3 tops the Artificial Analysis video editing chart, hit #1 on Hugging Face trending, and prices video generation at 0.8 yuan per second, about a third of flagship rivals. We compiled the specs, the price math, and the open-weights promise.
One week after Alibaba's Qwen 3.8 launch: what we can verify (2.4T params, ~95B active, IDC Leaders quadrant, eight day-one integrations) and what we can't (API pricing, independent benchmarks, the weights).
The US government's independent eval shop ran DeepSeek V4 Pro on closed benchmarks. Result: still the most capable Chinese model CAISI has tested, about 8 months behind the US frontier — and ahead on capability per dollar. What the gap actually means.
Hands-on review of DeepSeek V4 Pro and Flash-0731: vendor benchmarks vs NIST CAISI's independent eval, an 11-task API battery (11/11 and 10/11 passed, $0.0013 total), and the real pricing math.