Ledger on GitHub. Hub here. Branches elsewhere.
drop-2026-09-09-us-am
| claim | label | order (🟢 robust / ⚠️ sensitive) | numbers | URL |
|---|---|---|---|---|
| Qwen3.8-Max shipped as a 2.4T MoE with 95B active parameters and a 1M-token context. | 🟢 confirmed | 🟢 robust | 2.4T; 95B; 1M | https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen3-8-max?_p_lc=1 |
| Qwen3.8-2.4T-A95B has native 262,144-token context and extends to 1,010,000 tokens. | 🟢 confirmed | 🟢 robust | 262,144; 1,010,000 | https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B |
| Reuters reported Chinese open-weight models as cheaper and more customizable than the best US closed models, with strength in code generation. | 🟢 confirmed | 🟢 robust | — | https://www.reuters.com/technology/artificial-intelligence/american-ai-model-makers-smell-an-opportunity-2026-08-12/ |
| Hugging Face said 59% of 178 Chinese releases above 20B parameters in 2026 were Apache 2.0 and 22% MIT. | 🟢 confirmed | 🟢 robust | 178; 59%; 22% | https://huggingface.co/blog/state-of-open-models-summer-2026 |
| Hugging Face said China-origin open-weight models reached 41% of supply and over 10 billion cumulative downloads. | ⚠️ partial | ⚠️ sensitive | 41%; 10B+ | https://pandaily.com/china-open-source-llm-hugging-face-100-billion-downloads-aug2026 |
| DeepSeek V4-Pro-0813 was reported as a 1.7T-parameter GA/open-weights release under MIT. | ⚠️ partial | ⚠️ sensitive | 1.7T | https://aireiter.com/blog/deepseek-v4-pro-ga-api-guide |
| DeepSeek V4-Pro-0813 was also described as MIT-licensed open weights on Hugging Face with benchmark parity claims. | ⚠️ partial | ⚠️ sensitive | 1.7T | https://mr.technology/payloads/deepseek-v4-pro-0813-ga-mit-17t-august-2026 |
| Gemini 3.7 Flash was reported at 340.1 output tokens per second. | ⚠️ partial | ⚠️ sensitive | 340.1 tok/s | https://www.techzila.in/2026/08/gemini-3-7-flash-explained-features-pricing-benchmarks.html |
| Gemini 3.7 Flash was also described as a low-cost multimodal model with $0.75 input and $3.75 output per million tokens. | ⚠️ partial | ⚠️ sensitive | $0.75; $3.75 | https://www.dervity.com/blog/gemini-3-6-flash-3-5-flash-lite-guide |
| Qwen3.8-Max was described as text-image-video multimodal with a 1M headline context and day-one API availability. | ⚠️ partial | ⚠️ sensitive | 1M | https://www.digitalapplied.com/blog/qwen3-8-max-full-release-benchmarks-open-weights |
| Qwen3.8-Max pricing was reported as $2 input / $6 output per million tokens. | ⚠️ partial | ⚠️ sensitive | $2; $6 | https://www.developersdigest.tech/blog/qwen-3-8-max-release-2026 |
| DeepSeek V4-Flash-Vision-Exp was reported as a 305B multimodal open model with vLLM/SGLang serving recipes. | ⚠️ partial | ⚠️ sensitive | 305B | — |