The Hub is quieter than usual. Three named lab updates are live this morning. ThinkingBox, Kolibri-1 and PixelUMM are the notes.
On the Hub
Microsoft ThinkingBox: grade the database, not the tool calls
Microsoft and Hugging Face published ThinkingBox on 3 October. The benchmark runs 507 stateful business workflows twenty times from a clean backend. It grades agents on terminal database state and side effects, not just tool calls. Opus 5.5 leads with 67.16 percent pass at one, and Kimi-K3 is the strongest open model at 57.37 percent.1
Aleph-Alpha Kolibri-1: 78B German-English MoE
Aleph-Alpha lists Kolibri-1 with a release date of 3 October. The card shows a 78 billion parameter MoE with 3.46 billion active parameters, Apache 2.0, for German and English. The context window is one million tokens, with a 262k serving recommendation. The Hub entry appeared on 2 October and was updated on 3 October.2
NVIDIA PixelUMM: encoder-free pixels
NVIDIA lists PixelUMM on the Hub. It is an encoder-free unified image and video model that works in pixel space on a Qwen3-8B backbone. The card lists 15,199,672,064 parameters and a NVIDIA One-Way Noncommercial license for the checkpoint. The work is research only and the card appeared on 1 October.3
