dropkit.contents

Ledger on GitHub. Hub here. Branches elsewhere.

← Drops · 2026-09-14

drop-2026-09-14-kr-am

claimlabelordernumbersURL
Qwen3.8-Max open-weight base model was published in August 2026 and described as a 2.4T-parameter sparse MoE with about 95B active parameters.🟢 robustconfirm2.4T; 95Bhttps://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
Qwen3.8-Max’s managed version was described as having 1M context by default, with vision input and built-in tools in the official model notes.🟢 robustconfirm1Mhttps://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
Separate coverage said the Qwen3.8 open-weight release was the first Max-class open release, but the fully featured Max product still differed from the downloadable weights.⚠️ sensitiveconfirmfirst Max-classhttps://www.ideabosque.com/library/qwen3-8-open-weights-stripped-release-paywalling-the-frontier/
DeepSeek V4 Pro was released on 2026-08-13 and made available through API, with separate coverage also treating it as an open-weights release.🟢 robustconfirm2026-08-13https://api-docs.deepseek.com/news/news260813/
DeepSeek V4 Pro open weights were reported as 66 fp8 shards under MIT, and another source described the model at about 1.7T parameters with 893 GB weights.⚠️ sensitiveconfirm66 shards; MIT; 1.7T; 893 GBhttps://ofox.ai/blog/deepseek-v4-pro-0813-price-weights-benchmarks-api-access-2026/
Reuters reported DeepSeek V4 Pro pricing at $1.32 per million input tokens and $3.96 per million output tokens.🟢 robustconfirm$1.32/M in; $3.96/M outhttps://www.reuters.com/world/china/deepseek-releases-official-v4-pro-model-it-steps-up-expansion-2026-08-13/
A secondary report said DeepSeek V4 Pro was measured at 9.7% of theoretical peak on a fused MoE task versus Claude Opus 5 at 10.7%, framing the result as an efficiency claim.⚠️ sensitiveconfirm9.7%; 10.7%https://www.theneuron.ai/digest/everything-that-happened-in-ai-today-thursday-august-20-2026
OpenAI’s Jalapeño chip was reported to deliver 1.5x–1.9x more work per watt and 1.7x–3.6x lower end-to-end latency in benchmark coverage.⚠️ sensitiveconfirm1.5x–1.9x; 1.7x–3.6xhttps://www.theregister.com/systems/2026/08/25/openais-upcoming-jalapeno-chip-looks-like-itll-be-an-inference-beast/
OpenAI said Jalapeño would be deployed inside its infrastructure by the end of 2026.🟢 robustconfirmend of 2026https://www.cnbc.com/2026/08/26/openai-jalapeno-ai-chip-nvidia.html
OpenAI Frontier was described as an enterprise agent platform for building, deploying, and managing agents with shared context and integrations.⚠️ sensitiveconfirmenterprise agent platformhttps://www.cnbc.com/2026/08/26/openai-jalapeno-ai-chip-nvidia.html
A report said OpenAI tested Jalapeño against GPT-OSS 120B, DeepSeek R1, and Kimi K2.5 on public benchmarks.⚠️ sensitiveconfirmGPT-OSS 120B; DeepSeek R1; Kimi K2.5https://www.trendforce.com/news/2026/08/26/news-openai-debuts-jalapeno-ai-inference-chip-with-samsung-reportedly-supplying-hbm4/
Qwen3.8-Max was separately summarized as an August 2026 flagship open-weights-class model, corroborating the release narrative from another angle.⚠️ sensitiveconfirmAugust 2026https://the-decoder.com/alibabas-open-weight-qwen3-8-max-takes-on-long-horizon-ai-tasks-with-2-4-trillion-parameters/