Ledger on GitHub. Hub here. Branches elsewhere.
drop-2026-09-09-kr-am
| claim | label | order (🟢 robust / ⚠️ sensitive) | numbers | URL |
|---|---|---|---|---|
| DeepSeek-V4-Pro-0813 is the official release, superseding preview, with DSpark speculative decoding and agentic improvements. | 🟢 confirmed | 🟢 robust | 0 | https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813 |
| DeepSeek-V4-Pro-0813 was rolled out across app, web, and API on 2026-08-13. | 🟢 confirmed | 🟢 robust | 2026-08-13 | https://api-docs.deepseek.com/news/news260813/ |
| DeepSeek-V4-Pro-0813 pricing was reported at $1.32 per 1M input tokens and $3.96 per 1M output tokens, about 14× DeepSeek V4 Flash on output. | 🟢 confirmed | 🟢 robust | $1.32 / $3.96 per 1M | https://www.reuters.com/world/china/deepseek-releases-official-v4-pro-model-it-steps-up-expansion-2026-08-13/ |
| Artificial Analysis said DeepSeek V4 Pro 0813 scored 53 on its Intelligence Index, 8 points above April’s V4 Pro and 1 point above V4 Flash 0731. | 🟢 confirmed | 🟢 robust | 53 / +8 / +1 | https://x.com/ArtificialAnlys/status/2088440350734201149 |
| DeepSeek Harness shipped as a developer preview for agent-harness developers, with source code included. | 🟢 confirmed | 🟢 robust | v0.1 | https://www.deepseek.com/harness/en/ |
| DeepSeek Harness v0.1 was released in developer preview and open-sourced under MIT. | 🟢 confirmed | 🟢 robust | v0.1 / MIT | https://www.marktechpost.com/2026/08/17/deepseek-ai-releases-deepseek-harness-in-developer-preview/ |
| Qwen3.8-Max was officially released by Alibaba on 2026-08-03. | 🟢 confirmed | 🟢 robust | 2026-08-03 | https://www.alibabacloud.com/en/press-room/alibaba-unveils-qwen3-8-max?_p_lc=1 |
| Qwen3.8-Max was described as Alibaba’s most capable Qwen model to date. | 🟢 confirmed | 🟢 robust | — | https://qwen.ai/research |
| Qwen3.8-Max was described in secondary coverage as a 2.4T-parameter MoE with 95B active parameters and a 1M-token context window. | 🟡 partial | ⚠️ sensitive | 2.4T / 95B / 1M | https://promptailearning.com/ai-news/monthly/ai-news-recap-august-2026-every-story-that-mattered |
| Another roundup said Alibaba published Qwen3.8-2.4T-A95B with 2.4T parameters, 95B active, 262,144 native tokens, and a bespoke license. | 🟡 partial | ⚠️ sensitive | 2.4T / 95B / 262,144 | https://artifocial.com/blog/ai-news-2026-aug-10-16 |
| Alibaba’s Qwen3.8-27B weights went live on 2026-08-14 under Apache-2.0, with an official FP8 build. | 🟢 confirmed | 🟢 robust | 2026-08-14 | https://docs.tokios.com/models/qwen3.8 |
| Qwen3.8-27B is described as a dense 27B vision-language model aimed at local deployment. | 🟢 confirmed | 🟢 robust | 27B | https://www.aireleasetracker.com/model/qwen/qwen3.8-27b |
| A review said Qwen3.8-27B is a 27B Apache-2.0 vision-capable LLM from Alibaba’s Qwen lab. | 🟢 confirmed | 🟢 robust | 27B | https://simonwillison.net/2026/Aug/16/qwen-38-27b/ |
| Qwen3.8-27B was described in roundups as a dense 27B open-weights model with Apache 2.0 and 262,144 token context, extendable to 1M. | 🟡 partial | ⚠️ sensitive | 27B / 262,144 / 1M | https://www.webreactiva.com/blog/qwen-3-8-27b |
| Qwen3.8-27B reportedly showed 72.4 tok/s prose, 118.7 tok/s JSON, and 2,902 tok/s prefill on one RTX 5090. | 🔴 hype | ⚠️ sensitive |