# What Is Joey-1? Baseten’s Penny-a-Million Decision Model, Explained

URL: https://whatisagenticai.net/blog/what-is-joey-1/
Published: 2026-10-29
Publisher: Little Tough LLC, maker of AI Agent Assistant: Agentic AI (https://apps.apple.com/us/app/ai-agent-assistant-agentic-ai/id6761501260)

Joey-1 is Baseten’s 2B open-weights decision model, released October 29, 2026. It picks answers for $0.01 per million tokens. What it does and how to try it.

Joey-1 (written joey-1 by Baseten, and sometimes searched as "joey 1" or "Baseten Joey") is a tiny AI model that does not chat. You hand it a question and a few possible answers. It hands back the answer it picks and how sure it is.

Baseten released it on October 29, 2026, and the headline number is the price: one cent per million tokens. Here is what joey-1 is, who made it, what it costs, how to try it, and how it stacks up against the other "decision models" that arrived this month.

**TL;DR:** Joey-1 is a 2-billion-parameter open-weights decision model from Baseten’s research lab, Base Labs, released October 29, 2026. It is built on Alibaba’s Qwen3.5-2B and returns a choice plus a probability instead of writing text. On Baseten’s API it costs $0.01 per million input tokens and answers a typical request in about 20 milliseconds. The weights are free to download from Hugging Face under the MIT license.

## What is joey-1?

Joey-1 is the smallest member of Baseten’s decision model family. A decision model is a narrow AI that only answers closed questions. Think "Is this email spam?" or "Which team should get this ticket: billing, shipping or tech?" It does not write a reply. It gives one answer and a confidence score, in a single quick step.

Baseten says joey-1 handles three kinds of questions:

- **Yes or no.** It returns the chance the answer is yes.

- **Pick one.** It chooses from a list of named options and gives a probability for each.

- **Scores.** It rates something on an ordered scale, like 1 to 5 against a rubric.

It reads text and images. The training also included an "I don’t know" option, so the model can say it is unsure instead of guessing. Baseten pitches it as a first-line model: joey-1 takes the easy, high-volume calls, and its bigger siblings take over when it is not confident.

## Joey-1 at a glance

| Question | Answer |
| --- | --- |
| Who made it? | Base Labs, the in-house research lab of Baseten, an AI inference company |
| Released | October 29, 2026 |
| Size | 2B parameters (Baseten’s figure) |
| Built on | Alibaba’s open-weight Qwen3.5-2B, post-trained by Base Labs |
| What it returns | An answer plus a calibrated probability, not free text |
| Inputs | Text and images |
| API price | $0.01 per million input tokens |
| Speed | About 20 ms median for a 2,000-token request, per Baseten |
| Context | 16K tokens on Baseten’s shared API; Baseten says dedicated deployments go up to 262K |
| License | MIT, weights on Hugging Face as baselabs/joey-1 |

## Joey-1 vs elephant-1 vs mammoth-1

Joey-1 is the third model in a family. A week earlier, on October 21, Baseten shipped two bigger siblings. All three use the same training data and the same API, so a developer can switch between them by changing one line.

| Model | Size | Price per 1M tokens | Role |
| --- | --- | --- | --- |
| joey-1 | 2B | $0.01 | Fastest and cheapest, first line |
| elephant-1 | 9B | $0.035 | The balanced default |
| mammoth-1 | 27B | $0.10 | Most accurate |

Baseten gives a simple cost example. Sorting 10 million support tickets a month, at about 500 tokens each, is 5 billion tokens. That comes to about $50 a month on joey-1, versus $500 on mammoth-1. On speed, Baseten says joey-1 is about 2.5 times faster than elephant-1 (20 ms against 49 ms).

## Is joey-1 any good?

Baseten calls joey-1 the strongest decision model with 3 billion or fewer parameters. Baseten says those internal tests are "coming soon," so that claim cannot be checked yet.

The one public number is on the Jev Decision Index 0.4, a leaderboard for decision models. On its public split, the model card lists these scores:

| Model | Jev Decision Index 0.4 (public split) |
| --- | --- |
| elephant-1 (Baseten, 9B) | 72.24 |
| joey-1 (Baseten, 2B) | 63.30 |
| Cloudflare Clef | 61.71 |
| Jev 1.14 | 60.70 |
| Perplexity Decider Flash v1.2 | 60.06 |

That puts joey-1 at number 37 of 230 models and first among models under 4B, per Baseten. These are figures Baseten reported, so treat them like any launch-day benchmark. The fair reading: joey-1 trails its bigger sibling everywhere, but it beats several better-known decision models while costing a fraction as much.

## Joey-1 vs other decision models

October 2026 turned decision models into a crowded category. [Microsoft Decision-1](https://whatisagenticai.net/blog/what-is-microsoft-decision-1/) costs $0.042 per million input tokens. The [OpenAI Decisions API](https://whatisagenticai.net/blog/what-is-openai-decisions-api/) charges $0.10. [Perplexity Decider](https://whatisagenticai.net/blog/what-is-perplexity-decider/) is open-weight too and can run on a Mac. Joey-1’s angle is plain: the smallest and cheapest of the bunch, with an MIT license that lets companies run and change it however they like.

## How to try joey-1

- **Baseten Model APIs.** Developers call it through Baseten’s decision endpoint with the model ID decision-models/baseten-joey-2b-latest. Baseten says zero data retention is on by default.

- **Download the weights.** The model is on Hugging Face at baselabs/joey-1. The model card says self-hosting needs one CUDA GPU.

- **Fine-tune it.** Baseten says joey-1 can be post-trained on its Training Jobs product, and it published a cookbook for that.

There is no chat app. Joey-1 is a building block for software, not something you open on your phone.

## What this means for you

You will probably never call joey-1 yourself. But you will feel models like it. Every [AI agent](https://whatisagenticai.net/blog/how-ai-agents-work/) makes hundreds of tiny choices while it works: which tool next, did that step succeed, is this message safe. When those choices cost a cent per million tokens instead of dollars, agents get faster and cheaper to run.

The part a normal person cares about is the finished task, not the plumbing. That is the point of [agentic AI](https://whatisagenticai.net/what-is-agentic-ai/): expert agents that plan, check and finish real work. **Agentic AI** puts a team of up to four of them on your iPhone and iPad, with no account and no model shopping.

## Quick answers

**What is joey-1?** A 2B open-weights decision model from Baseten, released October 29, 2026. It answers yes-or-no, pick-one and scoring questions with a probability instead of writing text.

**Is joey-1 free?** The weights are free under the MIT license. Using it on Baseten’s API costs $0.01 per million input tokens.

**Is Baseten joey-1 open source?** It is open-weights under the MIT license, so you can download, run, change and use it commercially.

**What model is joey-1 based on?** Alibaba’s Qwen3.5-2B, post-trained by Base Labs with the same data and recipe as elephant-1 and mammoth-1.

**Can I chat with joey-1?** No. It only picks answers. For an AI that writes, researches and plans, use an assistant or an [AI agent app on iPhone](https://whatisagenticai.net/blog/ai-agents-on-iphone/).

New to these terms? Our [agentic AI glossary](https://whatisagenticai.net/agentic-ai-glossary/) explains decision models, inference and more.

**Sources:** [Baseten: Introducing joey-1](https://www.baseten.co/blog/decisions-for-a-penny-introducing-joey-1/) · [Hugging Face: baselabs/joey-1](https://huggingface.co/baselabs/joey-1) · [Baseten model library](https://www.baseten.co/library/joey-1/) · [The Next Web](https://thenextweb.com/news/baseten-joey-1-decision-model-open-weights) · [Unite.AI](https://www.unite.ai/baseten-releases-joey-1-a-2b-open-weights-decision-model/)