Emergent FM, for Thursday the 19th of November.
The big one. Cloudflare has released two open-weight models, Clairic and Clairic flash. Clairic pulls information out of complex documents, including images and PDFs. Flash is the lighter, faster one. Both run on Workers AI, and the weights are public. Cloudflare lists both at ten cents per million input tokens until December 1. If your apps use the Clairic preview endpoint, they move to the new Clairic model automatically as it rolls out.
Second. OpenAI has set January 11 as the end date for some GPT-5 fast tiers with EU data residency. Fine-tuned versions follow the same schedule. OpenAI recommends the same models on Standard processing. If you use those fast tiers, switch to the replacements.
Third, and this one you can try today. RadixArk has released DaisyRose, its first model, built on a Qwen model and trained with reinforcement learning. It lifts the Terminal-Bench pass rate from 5.6 percent to 12 percent. You can download it, or train your own agents with the Miles framework.
Fourth. Google has added Gemini 4 Neon preview to the Gemini API, on the paid tier. It costs three dollars per million input tokens.
Fifth. Tencent says it has opened Hy4 flash as an open-weight model. It reads text and images, and it writes text. It handles up to a million tokens of context.
Also today. LiveKit has added Neon, Google's Gemini preview model, to LiveKit Inference. Hugging Face has added a way for Transformers to stream MoE experts from disk. Cognition says merged pull requests grew 40 times while its team grew 2.7 times.
So, the one to act on. Check that your apps handle the new Clairic model.
That's the brief. The full stories are at emergent dot F M. You can sponsor this show at emergent dot F M, slash sponsors. This was read by an AI voice.