The Hub is about learning, training, and choosing this morning. ServiceNow posted AutoSynthData on Oct 2 to turn agent failures into synthetic training tasks.1 AllenAI opened Olmo-core 3 on Oct 1 for trillion scale MoE training.2 Cloudflare wrote about Clef and Clef-Flash on Oct 1, with Hub cards from Sep 30.3
On the Hub
AutoSynthData: enterprise agent training data
ServiceNow CoreAI posted AutoSynthData on the Hugging Face blog on Oct 2. The pipeline turns a target agent's failures into validated synthetic tasks for supervised fine tuning. Capability cards keep original prompts out of the generator. The demo runs on EnterpriseOps Gym. ServiceNow claims mean Pass@1 rises 7.2 points on a hybrid domain test and verifier success improves from 63.01 percent to 68.55 percent.1
Olmo-core 3: open trillion scale MoE trainer
AllenAI released Olmo-core 3 on Oct 1. It is an open MoE training stack aimed at trillion parameter models. The stack uses DDP expert parallelism to keep experts resident and route data to them. On 8x B300 GPUs a 47B MoE reaches about 52,000 tokens per second per GPU versus 19,400 before. The team reports 2.7x throughput and less than 5 percent drop when scaling experts from 8 to 128.2
Cloudflare Clef: open decision models
Cloudflare published a blog on Oct 1 about Clef and Clef-Flash. The models are Apache-2.0 decision models built on Qwen3.8-27B and Qwen3.5-9B. Hub cards were created on Sep 30. The models score schema choices in parallel instead of generating text. Cloudflare claims median latency of 38.8 ms for Clef-Flash and 209.3 ms for Clef on internal benchmarks, and the models are available on Workers AI and on Hugging Face.3
