Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents (Nov 23, 18:30 MSK): 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
ReleaseAPI / Product · ActionableTNW→ original

Baseten Releases joey-1: A Yes/No Model That Answers in 20 ms at $0.01 per Million Tokens

Baseten has released joey-1, an open 2-billion-parameter model that skips text generation entirely. Instead, it returns a yes/no answer or an option number along with a probability, for example "spam, 94%". A million tokens (pieces of text that pricing is calculated on) cost $0.01, and a request of 2,000 tokens is processed in 20 ms. That is enough for sorting support tickets, catching spam, and blocking harmful prompts without a large LLM. For routine classification, it makes more sense to hand the work to narrow models.

AI-processed from TNW; edited by Hamidun News
Baseten Releases joey-1: A Yes/No Model That Answers in 20 ms at $0.01 per Million Tokens
Source: TNW. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTNW ↗
Editor-in-Chief
Labeling 10 million tickets per monthabout $50 versus $500 on mammoth-1
Key metric #22.5 times faster than elephant-1; built on Qwen3.5-2B from Alibaba
Key metric #3Weights open on Hugging Face under the MIT license

Many companies run large language models on tasks that require no text at all: whether a message is spam, which queue a customer request should go to, whether a user is trying to get around a safeguard. A short answer of "yes", "no", or "option B" is enough for these. Baseten released joey-1, an open 2-billion-parameter model, precisely for this purpose. It does not write paragraphs. It returns an answer and a probability, for example "spam, confidence 94%". According to the company, it processes a typical request of 2,000 tokens (a token is a piece of a word, and pricing is calculated per token) in about 20 milliseconds, and a million tokens cost $0.01.

How joey-1 differs from a regular LLM

A regular language model responds with text, which then has to be parsed in code: searching for the word "yes," checking the format, dealing with extra chatter. Joey-1 works differently: its output is a ready choice plus a number showing how confident the model is in it.

The probability is useful in production. Confident cases can be handled automatically, while doubtful ones are sent to a human or a stronger model. This is a familiar pattern: a cheap filter first, then expensive analysis only for the contested cases.

Joey-1 belongs to one line of models with the larger mammoth-1 and elephant-1. They were released a week earlier by Baseten's research lab, Base Labs. A similar approach was shown by TypeSafe AI in September, so the "solver model" format is gradually becoming a separate niche.

What it costs and what is under the hood

Baseten offers a simple calculation: labeling 10 million support tickets per month on joey-1 would cost about $50. On mammoth-1, the same job would cost around $500. In terms of speed, joey-1 is roughly 2.5 times faster than elephant-1. For comparison, this month Microsoft released its own Qwen-based model at $0.042 per million tokens, which is about four times more expensive.

joey-1 is built on Qwen3.5-2B from Alibaba. Baseten fine-tuned it on 6.21 million examples of decisions. The weights are on Hugging Face under the MIT license: they can be freely used in commercial products and run on your own servers.

What to check before deployment

There is an important caveat. Baseten calls joey-1 the strongest decision model among those with no more than 3 billion parameters, but the tests behind that claim have not been published yet. On the model card, it ranks 37th out of 230 on the Jev Decision Index and shows the best result among models under 4 billion parameters. For a model of this size, the result is good, but an overall ranking does not guarantee that the model will handle your specific tickets or your definition of a "harmful" request. The speed and price figures are also currently provided by the company itself; the source contains no independent measurements.

The practical takeaway for teams is this. If a large LLM in your product is handling sorting, filtering, or routing, assemble a small set of your own labeled examples and compare joey-1 with your current solution on accuracy and cost. Given the open weights and the MIT license, a trial run costs almost nothing.

The second piece of advice is to use the probability as a threshold. Let confident answers be processed automatically, and send contested ones to manual review or to a more capable model. This way you can preserve quality and save money on the bulk of requests.

For Baseten itself, this is also a question of positioning. The company helps businesses run models, was valued at up to $13 billion in June, and is now releasing its own. Cheap decision models are a convenient way to attract customers who need high-volume inference rather than polished prose.

Story

Key Questions Answered

How joey-1 differs from a regular LLM?

A regular language model responds with text, which then has to be parsed in code: searching for the word "yes," checking the format, dealing with extra chatter. Joey-1 works differently: its output is a ready choice plus a number showing how confident the model is in it. The probability is useful in production. Confident cases can be handled automatically, while doubtful ones are sent to a human or a stronger model.

What it costs and what is under the hood?

Baseten offers a simple calculation: labeling 10 million support tickets per month on joey-1 would cost about $50. On mammoth-1, the same job would cost around $500. In terms of speed, joey-1 is roughly 2.5 times faster than elephant-1. For comparison, this month Microsoft released its own Qwen-based model at $0.042 per million tokens, which is about four times more expensive. joey-1 is built on Qwen3.5-2B from Alibaba. Baseten fine-tuned it on 6.21 million examples of decisions.

What to check before deployment?

There is an important caveat. Baseten calls joey-1 the strongest decision model among those with no more than 3 billion parameters, but the tests behind that claim have not been published yet. On the model card, it ranks 37th out of 230 on the Jev Decision Index and shows the best result among models under 4 billion parameters. For a model of this size, the result is good, but an overall ranking does not guarantee that the model will handle your specific tickets or your definition of a "harmful" request.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…