Baseten Introduces joey-1, a 2B Parameter Decision Model Priced at $0.01 per Million Tokens

30 October 2026 3 min read هدهد تك | Hudhud Tech

Baseten has introduced joey-1, a 2B parameter decision model developed by its research lab, Base Labs. Unlike traditional models, joey-1 returns a label along with a calibrated probability, targeting applications in high-volume routing, classification, and guardrail checks. The model is priced at $0.01 per million input tokens and is available with open weights on Hugging Face under the MIT license.

Key Features

Size: 2B parameters, post-trained from Qwen3.5-2B, with a maximum sequence length of 16,384 tokens (up to 262k on dedicated deployments, according to Baseten).
Deployment: Can be run on Baseten Model APIs or self-hosted on a single CUDA GPU (exact VRAM requirements not disclosed). No official quantized builds are available.
Performance: Ranked as the top model under 4B on the Jev Decision Index 0.4 public split, as reported by Baseten.
Best Score: Achieved a score of 63.30 on the Jev Decision Index 0.4, placing it at #37 out of 230 models.
Comparison: Trails its 9B counterpart, elephant-1, by 8.94 points on the same split (63.30 vs 72.24).
Summary: Offers 20 ms P50 decision-making for $0.01 per million tokens, but has less general knowledge compared to larger models.

Understanding joey-1

joey-1 is designed to answer typed questions in a single forward pass. Users provide a state (text or images) along with one or more questions, and the model returns probabilities for each option of every question. It does not generate free-form text that requires parsing by the application.

Supported Question Types:

Choice: Select one named option, such as identifying which team owns a ticket.
Yes/No (noul in the API): Returns the probability of a ‘yes’ response.
Score: Provides ordered levels, such as urgency, ranked from lowest to highest.

joey-1 is the third model in a series, following the recent releases of mammoth-1 (27B) and elephant-1 (9B). All three models share training data and the /v1/systemone API, allowing for easy switching with a single line of code.

Operational Mechanism of joey-1

Base Labs has post-trained Qwen3.5-2B with a complete fine-tuning of the language model, while keeping the vision encoder frozen. The model utilizes full-attention layers with a block-bidirectional mask, allowing all questions to be answered from a single shared-prefix prefill, minimizing the cost of additional questions.

Training Details from the model card:

Data: 6.21 million decision examples, comprising 58.3% core decision skills, 16.1% safety, 16.1% expert domains, and 9.5% public data.
Objective: Utilizes cross-entropy loss along with a Brier calibration term, including a trained “I don’t know” option.
Training Recipe: Conducted over 1 epoch with 43,390 steps, achieving a peak learning rate of 1e-5 using bf16 precision.
Hardware: Trained on 2 nodes with 8 H100 GPUs each.

The calibration feature allows users to set a confidence threshold, enabling uncertain answers to be escalated to larger models like elephant-1, mammoth-1, or even human intervention.