Ledger on GitHub. Hub here. Branches elsewhere.
drop-2026-09-08-kr-am
| claim | label | order | numbers | URL |
|---|---|---|---|---|
| DeepSeek-V4-Pro GA rolled out on app/web/API and the API model name stayed unchanged | 🟢 robust | confirmed | 2026-08-13 | https://api-docs.deepseek.com/news/news260813/ |
DeepSeek API docs map deepseek-v4-pro to DeepSeek-V4-Pro-0813, and deepseek-v4-flash to DeepSeek-V4-Flash-0731 | 🟢 robust | confirmed | 2026-08-25 | https://api-docs.deepseek.com/ |
| DeepSeek-V4-Pro supports the Responses API and tool calls, with a 1M-token context and 384K max output | 🟡 sensitive | partial | 1M; 384K | https://technode.com/2026/08/13/deepseek-v4-pro-api-update-adds-responses-api-support/ |
| DeepSeek V4-Pro architecture is described as Liquid MoE with spatial vision reasoning and NVIDIA H200 optimizations | 🟡 sensitive | partial | — | https://deepseek.ai/blog |
| DeepSeek V4-Pro uses three thinking effort levels: low, high, max | 🟢 robust | confirmed | low / high / max | https://api-docs.deepseek.com/updates/ |
| DeepSeek Harness is a TypeScript agent runtime with plugin architecture and breaking changes in the README | 🟢 robust | confirmed | breaking changes | https://github.com/deepseek-ai/deepseek-harness |
| Qwen3.8-Max is the latest flagship in the Qwen line and supports hybrid thinking by default | 🟢 robust | confirmed | hybrid thinking | https://docs.qwencloud.com/changelog/models |
| Qwen3.8-Max is described in QwenCloud docs as a sparse MoE model with 2.4T total parameters, about 95B activated per step, and a 1M-token context window | 🟢 robust | confirmed | 2.4T; ~95B; 1M | https://docs.qwencloud.com/changelog/models |
| Alibaba press coverage says Qwen3.8-Max API is available and weights were scheduled for release the following week | 🟢 robust | confirmed | next week | https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen3-8-max?_p_lc=1 |
Qwen3.8-Max open-source release Qwen/Qwen3.8-2.4T-A95B is documented as the open-source version of the flagship series | 🟢 robust | confirmed | 2 repos | https://docs.qwencloud.com/changelog/models |
| An independent arXiv benchmark paper reports DeepSeek V4 Pro 0813 at 62.0 on FlavourBench and Qwen 3.8 Max at 61.5 | 🟢 robust | confirmed | 62.0; 61.5 | https://arxiv.org/html/2608.20574v1 |
| DeepSeek V4-Pro’s benchmark uplift on DeepSWE is reported by vendor docs from 12.8 to 62.7 after GA | 🟡 sensitive | partial | 12.8 → 62.7 | https://api-docs.deepseek.com/updates/ |