Nous ships Arete-27B, an open Qwen fine-tune that cuts state framing
Supervised tuning cut propaganda-tagged answers from 22% to 0.9% with coding intact; plus free Vercel tracing and SageMaker's one-GPU-into-seven split.

Copy markdown
An open 27B that answers straight
Arete-27B is a Nous Research fine-tune of Alibaba's Qwen3.8-27B, Apache-2.0 and supervised-tuned across seven value categories to drop refusals and state-framed answers: propaganda-tagged responses fell from 22% to 0.9% and the fact score rose to 2.98/5, while coding held (Terminal-Bench-2.1 70% vs 67% base). Weights and 16 quants are on Hugging Face now, with no hosted provider yet, so you self-host it.
Vercel tracing is now free on every plan
Vercel Tracing hit GA across all plans, with span search by KQL in the dashboard or CLI and 14-day retention; new Next.js 16.5 projects get it with zero setup. If you've been bolting on third-party observability to debug AI routes and function spans, that's now built in.
SageMaker slices one GPU into seven
SageMaker real-time inference endpoints now support NVIDIA Multi-Instance GPU, carving one card into up to seven isolated slices so a small model stops hogging a whole GPU — AWS claims 18-34% lower cost per million output tokens. It's limited to P4d, P5, G7 and G7e, and excludes single-model and serverless endpoints.