Baseten opens joey-1, a 2B decision model at $0.01 per million input tokens

Source · Model card: mammoth-1 (27B) + elephant-1 (9B)

publicado (OK do Fabricio 30/10 BRT).

Baseten released joey-1 on 29 October 2026, the third open model from its Base Labs team after mammoth-1 and elephant-1 a week earlier. It is post-trained from Qwen3.5-2B with the same data and recipe as its siblings, takes text and images, and answers choice, yes/no and score questions from one shared prefill.

Where to get it. MIT weights on Hugging Face, and Baseten Model APIs at $0.01 per million input tokens with zero data retention by default. Baseten says that is a tenth of mammoth-1. Model APIs cap context at 16K; dedicated deployments go to 262K.

Vendor numbers (UnverifiedClaim). P50 end-to-end latency of 20 ms on a 2,000-token request, “5x faster than comparable decision models” and 2.5x faster than elephant-1 (49 ms). The card claims the best public-split score under 4B on Decision Index 0.4 (63.30) and 62.9 on a replica of the Artificial Analysis Decision Index.

What the boards show (30 Oct). The Decision Index 0.4.1 added joey-1 on 29 Oct: public 63.30, but Full 35.91 (#148 of 247), with ECE 0.044 and a 85 ms median on the board’s GPU. The private split, where Full comes from, is much harder on it than the public one, so the card’s ranking holds only for the public split. Not on JevBench or Artificial Analysis yet.

Score. The mammoth-1 page is scored on the 27B, so its scores do not change (7.0 / 8.1 / 12).