← All model news

Baseten's Base Labs releases joey-1, a 2B decision model that returns probabilities instead of text, with MIT-licensed weights

According to a MarkTechPost report dated October 29, 2026, Baseten's Base Labs released joey-1, a 2B decision model that returns a calibrated probability for each option in a single forward pass instead of free text. The weights are released under MIT, and the hosted API accepts up to 16,384 tokens. The report says it scored 63.30 on the public Jev split with a P50 latency of 20 milliseconds.

Published Updated

What changed

According to a MarkTechPost report dated October 29, 2026, Baseten's research team Base Labs released joey-1, a 2B-parameter decision model that does not generate free text; it takes a text or image state plus a set of typed questions and returns, in a single forward pass, a probability for each option of each question. The report says the weights are released on Hugging Face under the MIT license, and the model is also offered through Baseten Model APIs at the path /v1/systemone and is listed on OpenRouter. The article gives an API price of $0.01 per million input tokens, with output and cached input listed at $0. Those were the quoted prices on the day of the report; actual billing should be checked against the vendor's pricing page. joey-1 supports three question types: choice picks one of the given options; yes/no, whose field name in the API is noul, returns the probability of yes; and score returns ordered levels with the lowest first. The report says it was trained by full fine-tuning of Qwen3.5-2B, with the vision encoder frozen, block bidirectional masks in the full-attention layers, and a single shared prefix prefill across all questions, so the marginal cost of adding questions is small. On Model APIs the maximum sequence is 16,384 tokens; dedicated deployments are said to reach 262k. Self-hosting requires one CUDA GPU and a PyTorch 2.8 / CUDA 12.8 image; the report does not disclose VRAM requirements and does not list official quantized versions. The model card gives training details of 6.21 million decision samples, a cross-entropy loss plus a Brier calibration term, and an "I don't know" option that was trained. joey-1 is the third member of the series; Baseten released the 27B mammoth-1 and the 9B elephant-1 the previous week, and all three share training data and the same interface, with the report saying a switch takes one line of code. All benchmark numbers are Baseten's own: on the Jev Decision Index 0.4 public split, joey-1 scores 63.30, ranking 37th among 230 models, and Baseten calls it the highest score below 4B. elephant-1 scores 72.24, Cloudflare's Clef 61.71, and Jev 1.14 60.70. On Baseten's replication of the Artificial Analysis Decision Index, joey-1 scores 62.9 and mammoth-1 65.9. On latency, Baseten reports a P50 end-to-end latency of 20 milliseconds for an average 2,000-token request, versus 49 milliseconds for elephant-1, and the blog separately claims it is 5 times faster than comparable decision models without naming them.

What it means for developers

What this release changes is the shape of errors and the cost structure, not the ranking of another general-purpose model. High-frequency classification has usually been done by having a general LLM emit JSON that the application then parses, and the failure modes are format errors and hallucinated labels; a model like joey-1 constrains the output space to the given options, so parse failures largely disappear, and the remaining risk is that the probability judgment itself is inaccurate and that the threshold is set wrong. The trade-off is equally clear: a 2B decision model has weaker general knowledge than its larger siblings, and the report itself says it trails elephant-1 in every domain, so it fits first-line routing rather than replacing steps that need world knowledge or open-ended generation. The real use of calibrated probabilities is tiering. Since every option carries a probability and an abstain option was trained, engineering can escalate low-confidence samples to the 9B model, the 27B model or a human, concentrating expensive calls on hard samples; sharing one interface across the three models means the cost of building that escalation chain falls mainly on thresholds and evaluation rather than a code rewrite. But that also means adopters need two things they could previously skip: a labeled holdout set, and a policy that sets thresholds by cost. Without both, the price and latency advantages will not translate into better decisions in your own setting. Another easily overlooked variable is context. The hosted API truncates at 16,384 tokens, and 262k belongs only to dedicated deployments, which determines whether long tickets or long logs can take the hosted path; self-hosting needs only one CUDA GPU, but VRAM is undisclosed and there are no official quantized versions, so how long a sequence actually runs is something you have to verify yourself.

Integration and evaluation

The following is conditional analysis, not reported fact. If production today is prompt-based classification plus JSON parsing, start with a parallel shadow path rather than a direct replacement. The precondition is that the label set can be enumerated and that historical labels or human review data exist for a holdout set. Begin with the request class that has the highest volume, a closed label set and a low cost per mistake. The criterion is macro-averaged F1 and accuracy on your own holdout at a confidence threshold your business accepts, together with whether the parse failure rate goes to zero and whether the abstention rate stays within what humans can absorb; if accuracy does not beat the existing prompt baseline, or the abstention rate is too high, the conclusion should be not to switch. If the cost of a single mistake clearly exceeds the cost of one escalation call, treat the probability as a routing switch. Set a threshold, escalate samples below it to elephant-1, mammoth-1 or a human, and cap the escalation volume. The criterion is to sweep thresholds and find the lowest total cost of decision call cost plus escalation cost plus residual error cost, while also plotting a reliability curve to check whether samples the model scores at 0.8 are in fact correct about 80 percent of the time. The Brier calibration term in the report was tuned on its training distribution, and whether it remains calibrated after transfer to your domain can only be measured by you; this step cannot be skipped. If inputs routinely exceed 16k tokens or the data cannot leave your network, settle the deployment shape before discussing model quality. The 16,384 tokens of the hosted API, the 262k of dedicated deployment and the single GPU plus specified image required for self-hosting are three different paths. The criterion is to measure throughput and p99 latency on your target hardware and compare them with the latency of your existing pipeline; do not quote the vendor's platform-side 20 milliseconds, which measures their end-to-end environment.

Limits and open questions

The benchmark and latency figures in this article are all self-reported by Baseten and lack independent reproduction. The Jev Decision Index is a community leaderboard on Hugging Face, but the scores of the various models on it are likewise cited from Baseten's model card; the 20 millisecond and 49 millisecond comparison is limited to measurements taken within the same platform, and the blog's claim of being 5 times faster does not name what it compares against. MIT is the license stated on the weights page and is not a complete account of derivative rights in the base model or of the provenance of training data; check the license and data notes in the repository yourself before commercial use. joey-1 is not currently in KimiSeek's model catalog; the API availability and pricing mentioned in the report refer to supply on Baseten and OpenRouter. To check which models and billing are available on this platform, see the site's pricing page.

Three minutes to your first working call

Create an account and an API key. Calls use your account balance. 30% off your first top-up. Top up ¥500, pay ¥350, receive ¥500 in credit. Top up ¥1000, pay ¥850, receive ¥1000 in credit You can still start with a ¥10 trial top-up.

Sources and editorial process

Baseten 发布 joey-1:2B 决策模型,每百万输入 token 收费 0.01 美元

AI assisted the analysis and translation of the cited sources, followed by automated quality checks. Recommendations are not test results. Consult the current catalog for availability and pricing.