dropkit.contents

Ledger on GitHub. Hub here. Branches elsewhere.

← Drops · 2026-09-11

drop-2026-09-11-us-am

claimlabelordernumbersURL
Alibaba’s Qwen3.8-Max is its largest AI model yet, with open-weight positioning and aggressive pricing versus frontier peers.🟢 robust12.4T total; ~95B active; 1M context; $2/M input; $6/M outputhttps://www.reuters.com/business/retail-consumer/alibaba-unveils-its-most-capable-ai-model-date-not-far-behind-moonshots-size-2026-08-03/
Qwen3.8-Max’s open weights were later published as the Qwen3.8-2.4T-A95B checkpoint, with the max-class base model publicly downloadable.🟢 robust22.4T total; 95B active; published Aug 12–13; public checkpointhttps://arxiv.org/abs/2606.19348
DeepSeek V4-Flash is a very low-cost frontier model in benchmark economics, undercutting major closed peers.🟢 robust3~$0.14/M input; ~$0.28/M output; >100x cheaper in test-cost termshttps://www.reuters.com/business/retail-consumer/alibaba-unveils-its-most-capable-ai-model-date-not-far-behind-moonshots-size-2026-08-03/
DeepSeek-V4 technical report backs the family’s MoE shape and million-token context claims.🟢 robust4V4-Pro 1.6T / 49B active; V4-Flash 284B / 13B active; 1M context; 32T+ pretrain tokenshttps://arxiv.org/abs/2606.19348
DeepSeek V4 Flash-0731 was described as an MIT open-source weight release in secondary coverage.⚠️ sensitive5304B totalhttps://www.forkast.news/chinese-open-weight-frontier-compresses-five-labs-thirty-days-two-licensing-models/
Z.ai’s GLM-5.3-Flash was described as an open-weight frontier model trained on Chinese AI chips.⚠️ sensitive6320B total; 18B active; 1M contexthttps://www.requesty.ai/blog/open-weight-frontier-august-2026-glm-qwen-hy4
Tencent’s Hy4-preview was reported as a large open-source MoE with long-context serving and agent/tool-use relevance.⚠️ sensitive7770B total; 49B active; 1M contexthttps://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/
Hy4-preview’s open-weight release was paired with serving recipes and downloadable checkpoints, signaling infra-first deployment.⚠️ sensitive8BF16 + FP8; Apache 2.0; Tencent Cloud TokenHub/OpenRouter accesshttps://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/
Open-weight adoption is surging in China, with Chinese-origin models taking a much larger share of routed tokens.⚠️ sensitive9nearly half of OpenRouter tokens; up from ~11% a year earlierhttps://www.forbes.com/sites/drewbernstein/2026/08/03/chinese-ai-models-at-the-frontier/
Chinese open-model gap versus global frontier closed models is narrowing quickly.⚠️ sensitive102–3 months gap; previously 6–9 monthshttps://pandaily.com/china-open-source-llm-hugging-face-100-billion-downloads-aug2026
Meta and Nvidia also released open-weight models in mid-August, showing Silicon Valley is present in the open-weight race.⚠️ sensitive11mid-August 2026https://www.cnbc.com/2026/08/12/meta-nvidia-open-weight-ai-race-china.html
Open-weight release means weights are downloadable while training data and code remain private, limiting full inspection.🟢 robust12n/ahttps://www.reuters.com/technology/artificial-intelligence/