Baseten’s in-house research group, Base Labs, has released joey-1, a 2-billion parameter decision model priced at $0.01 per million input tokens. The open-weight model is available on Hugging Face under an MIT license and is designed to return probabilities and labels rather than free-form text, targeting high-volume enterprise tasks like content routing, classification, and safety guardrails.
According to the company, joey-1 fine-tunes Qwen3.5-2B using a block-bidirectional mask while keeping its vision encoder frozen. Baseten reports a P50 end-to-end latency of 20 ms for an average 2,000-token request on its platform, which it claims is roughly 2.5 times faster than its 9B elephant-1 model. On the Jev Decision Index 0.4 public benchmark, joey-1 scored 63.30, ranking top among models under 4 billion parameters.
The release provides developers with a structured, low-cost option for task automation that integrates into Baseten’s existing API setup. By offering deterministic classification at low price points, the model aims to reduce operational costs for high-throughput AI pipelines.
Why it matters
Enables ultra-low-cost, high-volume classification and routing at $0.01 per million tokens.
Provides open-weight MIT licensing, allowing custom deployment and fine-tuning without developer lock-in.
Sub-50ms latency allows seamless real-time guardrailing and request triage before routing to larger LLMs.
Source: marktechpost.com



