dropkit.contents

Ledger on GitHub. Hub here. Branches elsewhere.

← Drops · 2026-09-11

drop-2026-09-11-kr-am

claimlabelordernumbersURL
Alibaba officially launched Qwen3.8-Max in China as its largest and most capable Qwen model to date.🟢 robust🟢 robustAug 3, 2026https://www.alibabacloud.com/blog/alibaba-unveils-qwen3-8-max-its-largest-and-most-capable-flagship-model-to-date_603420
Qwen3.8-Max was described in third-party coverage as a 2.4T-parameter MoE flagship with about 95B active parameters and 1M-token context.⚠️ sensitive⚠️ sensitive2.4T total; 95B active; 1M contexthttps://www.japantimes.co.jp/business/2026/08/04/tech/china-ai-rival-us-model/
Alibaba Qwen3.8-Max is the first Qwen-Max-class model to open-source weights, and the announcement says the open weights would be released the next week.⚠️ sensitive⚠️ sensitivefirst Max-class open weightshttps://qwen.ai/research
DeepSeek-V4-Pro is generally available on app/web/API, with official docs and Reuters both confirming the launch and premium pricing shift.🟢 robust🟢 robust1.32 input / 3.96 output per million tokens; GA on Aug 13, 2026https://api-docs.deepseek.com/news/news260813/
DeepSeek’s V4-Pro-0813 is offered through app, web, and API, with stronger agent capabilities highlighted.🟢 robust🟢 robustAug 13, 2026; 1M contexthttps://www.reuters.com/world/china/deepseek-releases-official-v4-pro-model-it-steps-up-expansion-2026-08-13/
DeepSeek V4 Pro was positioned as a pricing and agent-capability move rather than a new architecture change.⚠️ sensitive⚠️ sensitive50% off-peak structure; 384k max output cited in coveragehttps://www.unite.ai/deepseek-ships-v4-pro-as-its-flagship-model-leaves-preview/
DeepSeek V4 Pro’s benchmark jump was reported as Terminal-Bench 2.1 from 72.1 to 87.9 and DeepSWE from 12.8 to 62.7.⚠️ sensitive⚠️ sensitive72.1→87.9; 12.8→62.7https://www.issuewire.com/15-frontier-ai-models-worth-watching-in-2026-ranked-and-benchmarked-via-aiccs-unified-api-1873938550876945
DeepSeek V4-Flash-Vision-Exp was described as a 305B multimodal model with vision encoding, MIT licensing, and vLLM/SGLang serving recipes.⚠️ sensitive⚠️ sensitive305B; MIThttps://www.buildfastwithai.com/blogs/ai-news-today-september-1-2026
GLM-5.3 is described in roundup coverage as an open-weights frontier model from Z.ai focused on agentic coding and cyber defense.⚠️ sensitive⚠️ sensitive753B total; 40B active; 1M contexthttps://fruition.net/frontier/
GLM-5.3 was described as an open-weights frontier model with 753B total parameters, 40B active, and 1M context.⚠️ sensitive⚠️ sensitive753B total; 40B active; 1M contexthttps://fruition.net/frontier/
Tencent’s Hy4-preview is mentioned as another Chinese frontier release with weights and competitive coding results.⚠️ sensitive⚠️ sensitiveno concrete numbers in firehosehttps://fruition.net/frontier/
TechCrunch cited a safety report saying GLM-5.2 was only a few months behind top U.S. models on cyber and bio risk dimensions.🟢 robust🟢 robustfew months behindhttps://techcrunch.com/2026/08/04/open-weight-ai-models-are-catching-up-to-the-frontier-the-safety-gap-remains/
Meta Muse Glimmer was reported as a 30B open-weight multimodal model with 128k context and 100+ languages.⚠️ sensitive⚠️ sensitive30B; 128k; 100+ languageshttps://promptailearning.com/ai-news/monthly/ai-news-recap-august-2026-every-story-that-mattered