Baseten Releases New AI Decision Model

The 2 billion parameter joey-1 model optimizes high-volume classification tasks at a fraction of the cost of generative models.

Updated on Oct. 29, 2026 in Artificial Intelligence

Isometric editorial illustration of a precision-machined metallic polyhedron, representing the structure of a specialized artificial intelligence decision model. — Baseten released joey-1, a 2-billion parameter decision model that optimizes high-volume classification tasks by outputting probability scores instead of generative text. AI Illustration. Upload story photo >

Live Poll

Would you consider using specialized, single-task AI models to automate your business's support workflows?

Baseten has released joey-1, a specialized 2 billion parameter AI decision model designed to output structured answers and probability scores rather than text. The model is currently available under the MIT license on Hugging Face.

Why it matters

By replacing generative text outputs with direct decision-making, the model accelerates and reduces the cost of high-volume tasks such as ticket routing and spam detection. This approach optimizes machine learning workflows that require binary or categorical classification over creative generation.

Joey-1 achieves 20-millisecond latency for 2,000-token requests at a cost of $0.01 per million tokens. The model is built on Alibaba's Qwen3.5-2B architecture and was trained on a dataset of 6.21 million decision examples.

The players

Baseten

A developer platform company focused on deploying and scaling machine learning infrastructure.

Alibaba

A global technology conglomerate whose Qwen open-weights models serve as the architecture foundation for various AI systems.

Hugging Face

An open-source hub that hosts and shares machine learning models, datasets, and demonstration apps.

The details

Joey-1 functions by mapping inputs directly to specific outputs and probability scores instead of employing the token-by-token generation characteristic of Large Language Models (LLMs). This architecture relies on its training set of 6.21 million decision examples to refine its categorization accuracy. By optimizing for classification tasks rather than natural language generation, the model maintains a compact 2 billion parameter footprint.

October 24, 2026: Baseten released joey-1.
September 2026: TypeSafe AI launched its Jev model.
October 2026: Microsoft released a decision model based on Qwen.

The Tech Race

Joey-1 competes in a burgeoning sector of decision-focused AI that prioritizes inference speed and classification accuracy over generative capability. It sits alongside recent releases from Microsoft and TypeSafe AI as firms shift from general-purpose LLMs toward task-specific architectures.

Developers can implement the model immediately via Hugging Face for tasks like automated ticket routing or spam filtering. Its $0.01 per million token pricing model provides a low-cost alternative for high-volume classification workflows.

The takeaway

The move toward specialized decision-based models suggests that industry workflows are increasingly pivoting away from general-purpose text generators. Watch for subsequent updates to the Jev Decision Index to see if joey-1 improves its ranking against established competitors.