📍 Home › Tools

Chinese AI Models Compared (2026): The Whole Board on One Page

Tools  Updated 2026-11-04 · 4 min read · #comparison #open-weights #export-controls

The two-sentence version

Chinese labs now ship models that trade punches with the western flagships at a fraction of the serving cost, and 2026’s twist is that half of them are open weights you can run on your own PC. Which one is right for you depends on one question: do you want to chat with a product, or own the model?

Status board (checked 2026-11-04 — every page below has the details)

AA Index = Artificial Analysis Intelligence Index v5, best reasoning setting, from leaderboard snapshots dated Nov 3–4. For scale, the top western models sit around 59–60 (Claude Opus 5.5, Gemini 4 Carbon).

Speeds on home hardware run from 4–6 tok/s (748B on a 12 GB RTX 5070) to 148 tok/s (GLM-5.3-Flash on three cards) — the whole math is on the pillar page.

Why “open weights” is the whole story

1. Price pressure. Flash-class models cost 1/10–1/40 of western flagships per token ($0.15/$0.50 per million for GLM-5.3-Flash, $0.15/$0.60 off-peak for DeepSeek’s deepseek-flash), which is why US startups started switching.
2. Sovereignty. A model you downloaded cannot be switched off by a headline — yours, your box.
3. The trade beat. China’s Oct-16 catalogue draft (Announcement No. 48; 40 technologies, 23 of them AI; comments until Nov 15) targets weights and core tech leaving China. Weights already published open-source are outside its scope, and calling a Chinese model through an API or app from abroad is not a technology export. Pre-export registration is already in force for new transfers: MiniMax (HKEX filing) and DeepSeek (per state media) said on Oct 23 they had completed it. Every status board on this site carries the date it was last true.

Pick in one line each

• Just want a free assistant → Qwen Studio (no paywall, 1M context) or DeepSeek for reasoning and code; Kimi for entire bookshelves of PDFs. Doubao is voice-first and zero friction — but only if you’re in mainland China, Hong Kong or Macau.
• Coding agents on a budget → GLM’s Coding Plan against Claude Code/OpenCode, or DeepSeek’s API.
• Own the model → start at run-100b-at-home, pick your engine (Strata or Project Maya), then the model page that matches your RAM.

FAQ

“Chinese AI” — is any of this legal to use commercially in the US/EU?

Today: open-weight licenses (MIT, Apache, community licenses) are the binding document. China’s draft rules exclude weights that are already public, and China’s commerce ministry says API, app and SaaS use from abroad is not a technology export. Nothing here is legal advice — check each model card, and the dated status boards for movement.

Which is smartest?

On the AA Intelligence Index (v5, Nov 3–4 snapshots): Kimi K3.1 Pro 46.5 leads the Chinese field, then Qwen3.8-Max 33.6, DeepSeek V4.1-Flash 32.6, GLM-5.3 31.7 and MiniMax M3.1-Flash 29.1 — all below the western leaders around 59–60. Rankings shuffle monthly; this page updates its date line when a flagship changes.

Why does this site cover the models AND the chat apps?

Same lab, two products: the app is how you try it, the weights are how you own it. Every row above links to both halves.