Baseten Releases Joey-1, a 2B Open-Weights Decision Model

Baseten on October 29, 2026 released joey-1, a 2B-parameter decision model developed by its in-house research lab Base Labs, priced at $0.01 per million tokens and available through Baseten Model APIs and as open weights on Hugging Face.

With joey-1, the Base Labs decision model family now spans three sizes. The announcement places the new model alongside mammoth-1 (27B) and elephant-1 (9B) from the October 21 family release. At that earlier release, Baseten described mammoth-1 as the top-ranked of 208 decision models on the Jev Decision Index and elephant-1 as the leading 9B decision model. All three models are trained on the same data mixture and share the same API, which the company says makes switching between them a one-line change. Baseten positions mammoth-1 for decisions where accuracy matters most, elephant-1 as the balanced default, and joey-1 for high-volume, latency-sensitive workloads.

Unlike generative LLMs, a decision model does not generate text. It returns a label and a confidence score for that label in one forward pass, with the scores across all possible answers summing to 1. Given a state — text, optionally with images — and a set of questions, joey-1 returns a calibrated probability for every option from one shared-prefix prefill. The model supports three question types: yes/no, choice among named options, and ordered scoring levels. Baseten’s earlier family post lists decision-model use cases including ticket triage and routing, spam and phishing detection, prompt guardrails and jailbreak detection, and routing requests between small and frontier LLMs.

Baseten said moving an LLM workload to joey-1 is worthwhile when decision volume is high enough that cost per call determines project economics, or when the decision is a prerequisite for another step and its latency is added to every response. As an example, classifying 10 million support tickets a month at roughly 500 tokens each totals 5 billion tokens, which the company prices at about $50 a month on joey-1 versus $500 on mammoth-1. Baseten describes joey-1 as a first-line model, with heavier models such as elephant-1 or mammoth-1 available as an escalation path when joey-1 is uncertain.

Reported Latency and Benchmark Results

Baseten reported that at production scale joey-1 serves an average 2,000-token request at a P50 end-to-end latency of 20 ms, about 2.5x faster than elephant-1 at 49 ms, and described the model as 5x faster than comparable decision models. The company said that in its internal testing, with details to come, joey-1 is the strongest decision model with 3B or fewer parameters.

The model card reports that joey-1 scored 63.30 on the Jev Decision Index 0.4 public split, placing 37th of 230 models and recording the highest score among models under 4B parameters on that split. The card’s public-split table lists elephant-1 at 72.24, Cloudflare Clef at 61.71, Jev 1.14 at 60.70, and Perplexity Decider Flash v1.2 at 60.06. According to the card, joey-1 scores below elephant-1 in every evaluated area and above Jev 1.14 in four of the five: language, tools and automation, retrieval and classification, and arts and human taste.

On Base Labs’ internal replica of the Artificial Analysis Decision Index, the card reports joey-1 at 62.9, behind mammoth-1 at 65.9, elephant-1 at 65.2, and d1-ultra at 65. Baseten’s model library page states that on release in October 2026, joey-1 is the top 2B model on the Jev Decision Index 0.4.

Training Data and Recipe

According to the model card, joey-1 was post-trained from Qwen3.5-2B using the same data mixture and recipe as elephant-1 and mammoth-1, run through Baseten Training Jobs. The process was a full fine-tune of the language model on text and image decisions with the vision encoder frozen, using a block-bidirectional attention mask on full-attention layers and a maximum sequence length of 16,384 tokens.

The training dataset comprised 6.21 million decision examples in a mixture the lab calls Mix 8: 58.3% core decision skills, 16.1% safety and abstention, 16.1% judging and expert domains, and 9.5% public decision data. The training objective combined cross-entropy with a Brier calibration term, with lambda set to 1, and included a trained “I don’t know” option for cases where the model is uncertain about its answer. The recipe ran one epoch of 43,390 steps at a peak learning rate of 1e-5 with the AdamW optimizer, beta values of 0.9 and 0.95 and no weight decay, a 3% warm-up, cosine decay, and bf16 precision across two nodes of eight H100 GPUs each, 16 GPUs in total.

Availability, Pricing, and Deployment

joey-1 is live on Baseten Model APIs through the /v1/systemone endpoint under the model ID decision-models/baseten-joey-2b-latest, with zero data retention on by default. Listed pricing is $0.01 per million input tokens, with cached input and output listed at $0.00. The family pricing table lists elephant-1 at $0.035 and mammoth-1 at $0.1 per million tokens, and the announcement describes joey-1’s price as one tenth of mammoth-1’s.

On Model APIs, the decision models support a 16k context window and are optimized for low-latency workloads. Dedicated deployments unlock the full 262k context window and support fine-tuned weights; Baseten said they can also deliver faster TTFT and TPS (time to first token and tokens per second) through workload-specific optimizations and co-location with LLM workloads. elephant-1 and joey-1 are available for post-training on Baseten Training Jobs, and the company has published a fine-tuning cookbook on GitHub.

The weights are open under the MIT license. For self-hosting, the model card’s setup script installs a pinned serving runtime, OneJev with a serving patch, and requires one CUDA GPU. The card notes that requests longer than the trained length limits are refused rather than truncated.