Baseten has released joey-1, a 2B decision model priced at $0.01 per million input tokens. The system returns a label and a calibrated probability rather than generated text, targeting high-volume routing, classification, and guardrail checks. Weights are available under the MIT license on Hugging Face.
Specs and performance
joey-1 contains 2B parameters and is post-trained from Qwen3.5-2B. It supports a maximum sequence length of 16,384 tokens on Model APIs, though Baseten claims 262k tokens are possible on dedicated deployments. The model runs on Baseten APIs or self-hosted on a single CUDA GPU, although specific VRAM requirements are not disclosed. No official quantized builds are listed.
On the Jev Decision Index 0.4 public split, joey-1 scores 63.30. This ranks it 37th out of 230 models and makes it the top performer under 4B parameters. It trails its 9B sibling, elephant-1, by 8.94 points on the same split (63.30 versus 72.24). The bottom line is a reported 20 ms P50 latency for decisions at the stated price point, though it possesses less general knowledge than its larger counterparts.
How it works
joey-1 answers typed questions in one forward pass. You send a state, such as text or images, and one or more questions. The model returns a probability for every option of every question. It does not write free-form text that your application must parse.
The system supports three question types:
- choice: Pick one named option, such as which team owns a ticket.
- yes/no: Returns the probability of yes. The API refers to this as
noul. - score: Ordered levels from lowest to highest, such as urgency.
joey-1 is the third member of a family. Baseten shipped mammoth-1 (27B) and elephant-1 (9B) last week. All three share training data and the /v1/systemone API, so switching between them is a one-line change.
Base Labs post-trained Qwen3.5-2B with a full fine-tune of the language model. The vision encoder stays frozen. Full-attention layers use a block-bidirectional mask. All questions are answered from one shared-prefix prefill, so extra questions add little cost.
Training details from the model card include the following:
- Data: 6.21M decision examples. The mix is 58.3% core decision skills, 16.1% safety, 16.1% judging and expert domains, and 9.5% public data.
- Objective: Cross-entropy plus a Brier calibration term, with a trained “I don’t know” option.
- Recipe: One epoch, 43,390 steps, peak learning rate 1e-5, bf16.
- Hardware: Two nodes of 8 H100 GPUs.
Calibration allows you to set a confidence threshold. Uncertain answers can escalate to elephant-1, mammoth-1, or a human.
Speed and cost
Baseten reports a P50 end-to-end latency of 20 ms for an average 2,000-token request. That is about 2.5x faster than elephant-1 (49 ms). The blog’s TL;DR also claims 5x faster speeds than comparable decision models, without naming them.
Model APIs charge $0.01 per 1M input tokens; output and cached input are listed at $0.00. elephant-1 costs $0.035 and mammoth-1 costs $0.10.
Baseten’s example calculates 10M support tickets a month at about 500 tokens each as 5B tokens. That is about $50 on joey-1, versus $500 on mammoth-1.
Benchmark comparison
All scores below are self-reported on Baseten’s model card.
On the Jev Decision Index 0.4 public split, joey-1 scores 63.30. That ranks #37 of 230 models and is the top score under 4B. The Jev Decision Index is a community leaderboard on Hugging Face.
On the same split, Cloudflare’s 27B Clef scores 61.71 and Jev 1.14 scores 60.70.
joey-1 trails elephant-1 in every area. It beats Jev 1.14 in 4 of 5 areas, including language, tools and retrieval.
On Baseten’s replica of the Artificial Analysis Decision Index, joey-1 scores 62.9, versus 65.9 for mammoth-1.
Comparison table
The following table compares joey-1 against other small decision models:
| Feature | joey-1 | elephant-1 | Perplexity Decider Flash v1.2 | Cloudflare Clef-Flash |
|---|---|---|---|---|
| Developer | Baseten (Base Labs) | Baseten (Base Labs) | Perplexity | Cloudflare |
| Parameters | 2B | 9B | 4B | 9B |
| Base model | Qwen3.5-2B | Qwen3.5-9B | Qwen3.5-4B | Qwen3.5-9B |
| License | MIT | MIT | Apache-2.0 | Apache-2.0 |
| Inputs | Text, images | Text, images | Text, images | Text, JSON, images, video |
| Max context | 16,384 (Model APIs); 262k on dedicated | 16k (Model APIs); 262k on dedicated | 262,144 (trained to about 32k) | 16,384 default |
| API price per 1M input tokens | $0.01 | $0.035 | $0.01 | Not disclosed |
| Jev Decision Index 0.4 (public)* | 63.30 | 72.24 | 60.06 | Not disclosed |
| Reported latency** | 20 ms P50 | 49 ms P50 | Not disclosed | 38.8 ms median |
| Self-host hardware | 1 CUDA GPU | Not disclosed | About 9 GiB BF16 weights on a CUDA GPU | Tested on 1 H200 |
| Weights access | Public | Public | Private repo, token required | Public |
*Jev 0.4 scores for all models come from Baseten’s joey-1 model card. **Latency figures are not comparable. Baseten measures end-to-end on its platform; Cloudflare reports its own Decision Index 0.2.1 run.
Deployment options
- Model APIs: Call
decision-models/baseten-joey-2b-latestat/v1/systemone. Zero data retention is on by default. - Dedicated deployments: Baseten says these unlock 262k context, fine-tuned weights and co-location with LLM workloads.
- Self-hosted: Download from Hugging Face and run
setup.sh. It needs 1 CUDA GPU and a PyTorch 2.8 / CUDA 12.8 image.
You can fine-tune joey-1 on Baseten Training Jobs using the fine-tuning cookbook. joey-1 is also listed on OpenRouter at $0.01 per 1M input tokens.
What it means
For teams managing queues or filtering content, joey-1 offers a specific tool: a fast, cheap classifier that returns a decision rather than text. You can route a ticket to a specific team or flag a safety issue without waiting for a long generation. The price is low enough to test on large volumes of data without breaking the budget. However, the model is not a general assistant. It lacks the broad knowledge of larger siblings and should be used where a simple, binary or multi-choice answer is required.