Baseten releases joey-1, a $0.01/M token open-weights decision model

What joey-1 does

Baseten’s joey-1 is a 2 billion parameter open-weights AI model designed for decision tasks, not text generation. Unlike traditional language models, it returns an answer (e.g., yes/no or multiple-choice) alongside a probability score, enabling faster and more interpretable outputs for specific applications.

The model is optimized for tasks like routing support tickets, detecting spam, or blocking harmful prompts. Baseten claims it answers a typical 2,000-token request in 20 milliseconds—2.5 times faster than its larger model, elephant-1. At $0.01 per million tokens, it is priced significantly lower than Microsoft’s comparable decision model ($0.042 per million tokens).

Performance and competition

Joey-1 ranks 37th out of 230 models on the Jev Decision Index, a benchmark managed by competitor TypeSafe AI. Baseten states this is the highest score for models with fewer than 4 billion parameters, though it has not released the underlying test data.

The model joins Baseten’s existing decision models, mammoth-1 and elephant-1, released last week by its research arm, Base Labs. All three models prioritize speed and cost efficiency over generative capabilities, targeting enterprise workflows that require rapid, deterministic outputs.

Microsoft recently launched its own decision model, built on Alibaba’s Qwen, with pricing at $0.042 per million tokens. TypeSafe AI’s Jev, released in September, operates on a similar principle to joey-1.

Availability and training

Joey-1’s weights are available for free on Hugging Face under an MIT licence, allowing developers to modify, deploy, or fine-tune the model without restrictions. It is built on Alibaba’s Qwen3.5-2B and trained on 6.21 million decision examples.

Baseten’s commitment to open weights contrasts with the proprietary approaches of competitors like Microsoft, potentially lowering adoption barriers for startups and enterprises.

Baseten’s recent moves

The release follows Baseten’s $1.5 billion funding round in June 2026, which valued the company at up to $13 billion. In September 2026, Baseten acquired Blaxel, a sandbox startup, signaling its expansion into niche AI tooling.

Baseten’s focus on decision models—rather than general-purpose AI—aligns with its strategy to carve out a specialized segment in enterprise AI infrastructure, competing directly with incumbents and startups alike.

What this means

Lazyfounder analysis — our interpretation, not reported fact.

Joey-1 reflects a growing shift toward specialized, cost-efficient AI models for enterprise tasks like routing support tickets or detecting spam. Unlike general-purpose models, decision models like joey-1 trade text generation for speed, accuracy, and transparency—delivering answers with probability scores. For founders and operators, this signals a viable alternative to high-cost, high-latency models, especially for tasks requiring rapid, deterministic outputs.

Baseten’s trajectory—from a $1.5B fundraise to acquiring Blaxel—suggests it is positioning itself as a contender in the AI infrastructure space, directly competing with players like Microsoft and TypeSafe AI. The MIT-licensed weights also lower the barrier to adoption, allowing startups to experiment without vendor lock-in.

Key takeaways

• Joey-1’s $0.01 per million tokens pricing is 4x cheaper than Microsoft’s comparable model, a critical advantage for high-volume tasks like support ticket routing.
• The model’s 20ms response time and probability-scored answers make it suited for real-time decision-making, a gap in today’s text-heavy AI landscape.
• Baseten’s $13B valuation and acquisition of Blaxel indicate aggressive scaling, but its claims about joey-1’s performance remain unverified without published benchmarks.
• Open-weights availability on Hugging Face (MIT licence) lets operators fine-tune or deploy the model without proprietary restrictions, unlike closed alternatives.

Related on Lazyfounder

• Microsoft
• More ai news
• Sarvam partners with India’s defence ministry for AI tech as Tata Group and Microsoft make headlines