News / Weekly Model Watch
Weekly
Model Watch
Auto snapshot
Updated 2026-09-06
Living page — updated weekly. Current data snapshot: week of 2026-08-20.
China Model Watch — Weekly Chinese AI Model Status
Registry snapshot from ChinaModelAPI’s flagship table (asOf 2026-08-20). This post is the automated weekly cadence for SEO/GEO freshness — it does not invent new model versions.
Direct answer
Current Chinese stack we track for API builders: Qwen3.8-Max-Preview · DeepSeek-V4-Pro · DeepSeek-V4-Flash · GLM-5.2 · Kimi K3…. Western comparison baselines remain GPT-5.6 Sol and Claude Opus 5. Production video default is still Seedance 2.0 while 2.5 sits in an August window unless GA is confirmed.
Chinese models (tracked)
| Model | Maker | Notes |
|---|---|---|
| Qwen3.8-Max-Preview | Alibaba | 1M |
| DeepSeek-V4-Pro-0813 | DeepSeek | 正式版 (GA) 2026-08-13; enhanced agents; model id deepseek-v4-pro unchanged; minor API price adjustment Harness v0.1 (dsh) open-sourced 2026-08-13 evening: MIT dev-preview agent harness, everything-is-a-plugin on Cordis; minimal mode = official benchmark framework. Harness multimodal (as of 08-20): rc.7 (08-17) durable image attachments MCP/ACP; rc.8 (08-19) configurable native image requests for DeepSeek adapters + image input for /goal //plan; npm latest=rc.7, rc.8 on next; V4-Pro/V4-Flash remain text-only (official API changelog unchanged since 08-13). |
| DeepSeek-V4-Flash | DeepSeek | 1M |
| GLM-5.2 (prod) / GLM-5.3 (released 08-14) | Zhipu/Z.ai | GLM-5.3 update (as of 08-19): API pricing published on docs.z.ai international USD — $1.40 input / $0.26 cached / $4.40 output per 1M, same as 5.2; OpenRouter listed same price 08-18. First independent eval: Artificial Analysis II v4.1.1 = 60, rank 8/182 (max-effort 59.5 vs Kimi K3 59.7 / GPT-5.6 Sol 60.9 / Opus 5 61.5), flagged 'very verbose' (170M vs 72M median tokens on index run). Weights NOT yet on HF zai-org (checked 08-19), expected ~08-28 after safety hardening. Vendor numbers: Terminal-Bench 3.0 4.6→28.3, DeepSWE 46.2→66.9, CyberGym 84.5%, 2,436 real vulns (1,097 critical/high per chart). API breaking change: thinking.type 'disabled' removed. English press: VentureBeat (Cursor vuln claim, unconfirmed), The Decoder, Interconnects (~750B params, ~$1B ARR reportedly), Reuters (trusted access), HN 1,167-pt thread. Production stays glm-5.2 until 5.3 GA and healthy through relays. |
| Kimi K3 | Moonshot | 1M |
| Seedance 2.0 (prod) / 2.5 (August window) | ByteDance | 2.5: ~30s native; early 年框 access; GA delayed from early/mid-July toward August |
| HappyHorse | Alibaba | video |
Western baselines
| Model | Maker | Notes |
|---|---|---|
| GPT-5.6 Sol | OpenAI | Flagship Sol tier; Terra/Luna cheaper |
| Claude Opus 5 | Anthropic | Opus line; Sonnet 5 mid tier |
| Grok 4.6 | xAI | Listed in xAI official docs as recommended flagship; 500k context; $2/$6 per 1M (<200k prompt); knowledge cutoff 2026-02-01. Western baseline only — not routed by ChinaModelAPI. |
| Gemini 3.7 Flash | Announced 2026-08-13 (3 weeks after 3.6 Flash); 'most intelligent workhorse' for coding+agents; gemini-3.7-flash stable in Gemini API docs; 1,048,576 input / 65,536 output; thinking low/med/high; vendor benchmarks DeepSWE 65.3% / FrontierCode 43.6%; intro $0.75/$3.75 per 1M through 2026-12-31, then $1.50/$7.50. Western baseline only — not routed by ChinaModelAPI. |
How this automation works
- Weekly (Mon) — auto publish this snapshot from
data/model-watch/current-flagships.json. - Deep briefs — written when flagships change or when a research session fills
draft-*.md. - Sources policy — official docs first; X only for discovery. This auto post does not scrape social feeds.