Baseten's Base Labs turns Qwen into decision models with 6.21M examples

Baseten's research lab Base Labs, led by Max Kirkby, released two open-weight decision models on October 21st: the 27-billion-parameter mammoth-1 and the 9-billion-parameter elephant-1, both built by post-training Qwen models to return calibrated probabilities instead of generated answers. Base Labs says mammoth-1 leads the Jev Decision Index and elephant-1 ranks among the leaders at its smaller size, and it published enough of the method, trained on 6.21 million examples, for the benchmark claims to be tested.

Baseten's Base Labs turns Qwen into decision models with 6.21M examples Led by Max Kirkby, the lab reports top benchmark scores for two open-weight models built to return calibrated probabilities instead of generated answers. By Ryan Merket https://runtimewire.com/author/ryan-merket · Published Primary source: X https://x.com/baselabs/status/2112955587059450035 Why it matters Base Labs is showing how a post-training team can turn an existing open model into a specialized decision system, and it has published enough of the method to make the benchmark claim testable. Whether those gains transfer to real workflows will matter more than the leaderboard position. Max Kirkby https://x.com/maxkirkby?ref=runtimewire and Base Labs, the research lab inside Baseten, released mammoth-1 https://runtimewire.com/models/huggingface/baselabs-mammoth-1-fa430c248226c108 and elephant-1 https://runtimewire.com/models/huggingface/baselabs-elephant-1-c0d83b56ba3b9891 on October 21st, turning Qwen models into specialist systems that score choices rather than generate prose. The lab says its 27-billion-parameter mammoth-1 leads the Jev Decision Index, while its 9-billion-parameter elephant-1 ranks among the leaders at a smaller size.… When you purchase through links in our articles, we may earn a small commission https://runtimewire.com/affiliate-monetization-standards . This doesn’t affect our editorial independence.