Skip to content
Neural Daily

ModelsNews

joey-1: a 2B open-weight decision model at 20 ms

🇭🇺 Olvasd el magyarul →

Baseten's research lab Base Labs released joey-1, a 2-billion-parameter decision model. Instead of text, it returns a single label with a calibrated probability, at 20 ms P50 latency and one cent per million input tokens. The weights are open under the MIT license.

By: Mikola Máté Patrik · · 2 min read

Baseten announcement card for the joey-1 decision model
Image: Baseten · Press kit · Official social image from the Baseten blog post · Source

What is a decision model?

Many LLM tasks don't need text, they need a decision: where should this ticket go, should this message be flagged, should this request be approved. A decision model skips generation and returns a label with a calibrated probability in a single forward pass. That makes it much cheaper and faster, especially when the decision sits inside a longer chain (Baseten).

joey-1 by the numbers

  • 2 billion parameters, post-trained from Qwen3.5-2B with the same data and recipe as mammoth-1 (27B) and elephant-1 (9B), released eight days earlier.
  • P50 latency of 20 ms for an average 2,000-token request, about 2.5x faster than elephant-1 at 49 ms.
  • Price on Baseten Model APIs: $0.01 per million input tokens, a tenth of mammoth-1.
  • Text and image input. 16k context on Model APIs, 262k on dedicated deployments.
  • Zero data retention by default, with MIT-licensed weights on Hugging Face.

According to the model card, joey-1 scores 63.30 on the public split of the Jev Decision Index 0.4, the best result under 4B parameters and 37th of 230 models (Hugging Face). For comparison, elephant-1 scores 72.24, Cloudflare Clef 61.71 and Jev 1.14 60.70. On Baseten's replica of the Artificial Analysis Decision Index it scores 62.9.

It handles three question types: choice picks from named options, noul answers yes or no, and score rates on ordered levels. A single request can carry several questions sharing one prefill. Training used 6.21 million decision examples on 16 H100 GPUs, with a Brier calibration term in the loss and a trained "I don't know" option.

What it means for you

Baseten's own example shows the scale: classifying 10 million support tickets a month at about 500 tokens each is 5 billion tokens, roughly $50 a month on joey-1 versus $500 on mammoth-1. If you run routing, moderation or guardrails at high volume, joey-1 can be the first-line filter, with uncertain cases escalated to a bigger model. All three models share one API, so switching is a one-line change. With open weights, you can also run and fine-tune it on your own infrastructure.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles