Baseten Releases joey-1: A 2B Decision Model Priced at $0.01 per 1M Tokens

Baseten has released joey-1, a 2B model from its in-house research lab, Base Labs. Rather than generating free-form text, the joey-1 decision model returns a label paired with a calibrated probability, making it purpose-built for high-volume routing, classification, and guardrail checks. Baseten prices it at $0.01 per million input tokens, and the open weights are available on Hugging Face under the MIT license.

Against the backdrop of 2026's crowded small-model landscape—where sub-4B models increasingly handle agentic subtasks and deterministic decisioning rather than open-ended generation—joey-1 positions itself as a low-cost, purpose-specific primitive for production pipelines.

Key Takeaways

• Size: 2B parameters, post-trained from Qwen3.5-2B. Maximum sequence length of 16,384 tokens (Baseten cites 262k on dedicated deployments).
• Runs on: Baseten Model APIs, or self-hosted on a single CUDA GPU (VRAM requirements not disclosed). No official quantized builds are listed.
• Performance: Baseten reports it as the top model under 4B parameters on the Jev Decision Index 0.4 public split.
• Best result: 63.30 on the Jev Decision Index 0.4 (public split), ranking #37 of 230 models.
• Weakest point: It trails its 9B sibling, elephant-1, by 8.94 points on the same split (63.30 vs. 72.24).
• Bottom line: Best case—20 ms P50 decisions at $0.01 per 1M tokens. Worst case—less general knowledge than its larger siblings.

What Is joey-1?

According to Baseten, joey-1 is a compact decision model designed to output structured labels and calibrated probabilities instead of prose. That design choice suits workloads where latency and cost dominate—such as request routing, content classification, and policy or safety guardrail checks—rather than tasks requiring broad world knowledge or long-form reasoning.

Derived from Qwen3.5-2B through post-training, the model carries a 16,384-token maximum sequence length in its standard configuration, with Baseten noting support for up to 262k tokens on dedicated deployments. It is accessible through Baseten's Model APIs or can be self-hosted on a single CUDA GPU. As of release, no official quantized builds have been listed.

Performance and Positioning

On the Jev Decision Index 0.4 public split, joey-1 scores 63.30, placing it #37 among 230 models and making it, per Baseten, the strongest model in the sub-4B class. The trade-off is clear when compared with its 9B sibling, elephant-1, which scores 72.24—an 8.94-point gap that reflects joey-1's reduced general knowledge capacity. For decision-oriented workloads, however, the smaller footprint delivers on the metrics that matter most: Baseten reports 20 ms P50 decisions at a price point of $0.01 per 1M tokens.

Availability

• Licensing: Open weights under MIT on Hugging Face (baselabs/joey-1).
• Deployment: Baseten Model APIs or self-hosted on a single CUDA GPU.
• Pricing: $0.01 per million input tokens via Baseten's API.

via MarkTechPost