OfficialVercel AI Gateway adds DeepSeek V4.1 Flash 1017 Ultrafast at over 1,000 tokens per second
Vercel published this on Nov 7, 2026 (source: vercel.com).
Vercel added the DeepSeek V4.1 Flash 1017 Ultrafast variant to AI Gateway in public preview, served through Fireworks' accelerated serverless endpoint. Vercel says the model generates upwards of 1,000 tokens per second with a 1M-token context window, no waitroom or session time limit, and provider pricing passed through with no gateway platform fee.
+1 sources
- Vercel AI Gateway model page · Official
OfficialRohan Varma, OpenAI Codex PM who earlier worked on Cursor, says he has left OpenAI
@TheRohanVarma on X published this on Nov 7, 2026 (source: x.com).
Rohan Varma announced on X that he has left OpenAI after contributing to Codex, thanking Codex leads Alexander Embiricos and Thibault Sottiaux. He was a product manager on the Codex team after previously working on Cursor; a next role was not announced.
OfficialShizuku AI Labs launches Shizuku-Duplex, a full-duplex audio synthesis model for AI characters
@ShizukuAILabs on X published this on Nov 7, 2026 (source: x.com).
Shizuku AI Labs announced Shizuku-Duplex on its official X account, calling it the first full-duplex audio synthesis model built for AI characters: it responds with natural backchannels while the user is still talking. The launch post includes a demo video; the company's site has no model card or documentation page yet.
OfficialQwen Code 0.27.0 adds parallel child questions in managed-agent sessions; SDK TypeScript 0.1.30 bundles it
Qwen Code published this on Nov 7, 2026 (source: github.com).
Qwen Code 0.27.0 lets managed-agent sessions hold several child questions at once and answer them in any order, adds a Web Shell task Q&A heatmap with an Ask-for-details side task, and brings repository filter allowlists and blocklists to the GitHub channel editor. The TypeScript SDK 0.1.30 release bundles this CLI version.
+1 sources
- Qwen Code SDK TypeScript 0.1.30 · Official
OfficialNous Research releases Arete-27B, a Qwen3.8-27B post-training it says cuts propaganda answers from 22% to about 1%
Nous Research published this on Nov 6, 2026 (source: nousresearch.com).
Nous Research released Arete-27B, an Apache-2.0 post-training of Alibaba's Qwen3.8-27B aimed at reinforcing Western values. On Nous's own 803-prompt evaluation across seven value categories, answers containing propaganda fell from 22% to about 1% and the fact score rose from 2.32 to 2.98 out of 5 with thinking off; the weights are on Hugging Face. The headline numbers come from Nous's internal evaluation and have no independent verification yet.
+2 sources
- Nous Research (@NousResearch) on X · Official
- Arete-27B weights on Hugging Face · Official
Officialllama.cpp merges support for Microsoft's Mage-VL-Cowork vision-language model
Llama.cpp published this on Nov 7, 2026 (source: github.com).
llama.cpp merged a pull request adding Mage-VL-Cowork to its mtmd multimodal support. Microsoft published the Apache-2.0 image-text-to-text model on Hugging Face on October 30 as a finetune of its Mage-VL; the contributor notes RoPE handling still needs some work but says the model runs fine.