Baseten Launches a 2B Parameter Model: No Chat, Only Probability Output, Delivers a Judgment in 20 Milliseconds
It returns calibrated probabilities, with a P50 latency of 20ms, and costs only $0.01 per million input tokens. This 2B parameter open-weight model shifts its core task from generating a paragraph of text to outputting a clear judgment result.
Baseten has released joey-1, a 2-billion-parameter open-weight decision model. It does not generate text, and its output is a calibrated probability. It delivers a P50 latency of 20ms, with an input cost of $0.01 per million tokens. This information comes from a report by MarkTechPost author Asif Razzaq.
We are long accustomed to letting models generate text: ask it a question, and it replies with a full paragraph; feed it reference materials, and it writes a summary. Joey-1 takes a completely different path: it only outputs a probability value. For example, an output of 0.82 means the model judges there is an 82% chance that the target event is a positive case. The value of a calibrated probability is that this 0.82 value stays as close as possible to the real occurrence frequency: among similar samples, roughly 82% of them are indeed positive cases. After business teams get this numerical value, they can make decisions directly based on preset thresholds, without needing another model to parse natural language outputs.
This type of model solves the pain points of high-frequency judgment scenarios. For links like content moderation, risk scoring, work order routing, recommendation filtering, and quality inspection sampling, most teams only need a fast, comparable score to trigger subsequent processing per preset rules. General-purpose chat models can also handle these tasks, but they carry extra decoding overhead for token generation, and their outputs also come with unstable formatting fluctuations. Joey-1 compresses output to a single probability, creating a shorter inference pipeline with far lower costs. Its 20ms P50 latency and price of $0.01 per million input tokens make it perfect for deployment at positions requiring repeated frequent calls.
Open-weight means the full model weights are accessible to users. Details including the exact license, context window length, and training data are not disclosed in the public materials, so users need to verify these details by themselves before deployment. The 2B parameter scale also makes it clear that this model is not designed to cover all tasks. It acts more like a dedicated judgment processor, suitable for scenarios with clear boundaries and stably defined labels. Probability calibration still needs to be validated on real business data distribution, and corresponding thresholds need to be readjusted when data distribution drift occurs.
One notable observation: the AI circle has been competing for better chat performance for a long time recently, while models like joey-1 go the opposite way: stripping away all natural language generation functions, leaving only the judgment capability. It is not necessarily smarter than general chat models, but it may be far closer to what many businesses actually want: a low-cost, fast score that can be directly mapped to actionable thresholds. Original article link: https://www.marktechpost.com/2026/10/29/baseten-joey-1-2b-decision-model/
发布时间: 2026-10-30 05:20