Thinking Machines and Baidu expand open-weight model options

Thinking Machines Lab has introduced Inkling 2 Small, a mixture-of-experts model with 594B parameters and 17B active during inference. The company plans to release its weights so businesses can adapt the model to their needs. Its performance is described as competitive with leading Chinese open-weight models. Baidu has released ERNIE 5.5 Flash Base, a pretrained mixture-of-experts model with 280B parameters and 14B active during token processing. Its weights are available on Hugging Face, but its license excludes use in Korea and Europe.

Yejin has developed the Korean-focused foundation model 예진 8B v3, building its tokenizer and conducting pretraining independently rather than adapting an existing open model. Training used Nvidia H200 GPUs provided through a government-backed computing support program. NousResearch’s Arete-27B, a text-generation model with 27.4B parameters, entered Hugging Face’s trending listings. Viggle also published Viggle-Remake, a video-to-video model, on the platform.

For inference infrastructure, Buoyant released Kinference, an Apache 2.0 control plane for Kubernetes and vLLM. It provides declarative routing policies, KV-cache-aware load balancing and per-tenant metrics. Buoyant reported a 95% reduction in median time to first token in its own benchmark, making the result a company-reported measurement rather than an independently validated performance claim.