Baseten 发布 2B 决策模型 joey-1，每百万 token 定价 0.01 美元

Add as a preferredsource on Google

Baseten has released joey-1, a 2B model from its in-house research lab, Base Labs. The joey-1 decision model returns a label and a calibrated probability instead of generated text. It targets high-volume routing, classification and guardrail checks. Baseten prices it at $0.01 per million input tokens, with open weights on Hugging Face under MIT.

TL;DR

Size: 2B parameters, post-trained from Qwen3.5-2B. 16,384-token max sequence (Baseten says 262k on dedicated deployments).

Runs on: Baseten Model APIs, or self-hosted on 1 CUDA GPU (VRAM not disclosed). No official quantized builds listed.

Performance: Top model under 4B on the Jev Decision Index 0.4 public split, per Baseten.

Best: 63.30 on Jev Decision Index 0.4 (public split), ranked #37 of 230 models.

Worst: Trails its 9B sibling elephant-1 by 8.94 points on the same split (63.30 vs 72.24).

Bottom line: Best: 20 ms P50 decisions for $0.01 per 1M tokens. Worst: less general knowledge than its larger siblings.

What is joey-1?

joey-1 is a 2B decision model from Baseten’s Base Labs that answers typed questions in 1 forward pass. You send a state (text or images) and 1 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.

TEXTTABLELAYOUTIs your PDF parser actually getting it right?omni_parse_bench · scores text, tables & layout · open sourceStar on GitHub→Sponsored

It supports 3 question types:

choice: pick 1 named option, such as which team owns a ticket.

yes/no (noul in the API): returns the probability of yes.

score: ordered levels, lowest first, such as urgency.

joey-1 is the third member of a family. Baseten shipped mammoth-1 (27B) and elephant-1 (9B) last week. All 3 share training data and the /v1/systemone API, so switching is a 1-line change.

How does joey-1 work?

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 1 shared-prefix prefill, so extra questions add little cost.

Training details from the model card:

Data: 6.21M decision examples. Mix: 58.3% core decision skills, 16.1% safety, 16.1% judging and expert domains, 9.5% public data.

Objective: cross-entropy plus a Brier calibration term, with a trained “I don’t know” option.

Recipe: 1 epoch, 43,390 steps, peak learning rate 1e-5, bf16.

Hardware: 2 nodes of 8 H100 GPUs.

Calibration lets you set a confidence threshold. Uncertain answers can escalate to elephant-1, mammoth-1 or a human.

How fast and cheap is joey-1?

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 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: 10M support tickets a month at about 500 tokens each is 5B tokens. That is about $50 on joey-1, versus $500 on mammoth-1.

How does joey-1 score on benchmarks?

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.

joey-1 vs. other small decision models

Featurejoey-1elephant-1Perplexity Decider Flash v1.2Cloudflare Clef-Flash

Parameters2B9B4B9B

Base modelQwen3.5-2BQwen3.5-9BQwen3.5-4BQwen3.5-9B

LicenseMITMITApache-2.0Apache-2.0

InputsText, imagesText, imagesText, imagesText, JSON, images, video

Max context16,384 (Model APIs); 262k on dedicated16k (Model APIs); 262k on dedicated262,144 (trained to about 32k)16,384 default

API price per 1M input tokens$0.01$0.035$0.01Not disclosed

Jev Decision Index 0.4 (public)*63.3072.2460.06Not disclosed

Reported latency**20 ms P5049 ms P50Not disclosed38.8 ms median

Self-host hardware1 CUDA GPUNot disclosedAbout 9 GiB BF16 weights on a CUDA GPUTested on 1 H200

Weights accessPublicPublicPrivate repo, token requiredPublic

*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.

How do you deploy joey-1?

Model APIs: call decision-models/baseten-joey-2b-latest at /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.

Key Takeaways

joey-1 returns labels plus calibrated probabilities, not text.

Baseten reports 20 ms P50 latency at $0.01 per 1M input tokens.

63.30 on Jev Decision Index 0.4 public: top model under 4B.

Trails elephant-1 by 8.94 points; best as a first-line router.

MIT weights on Hugging Face; self-host on 1 CUDA GPU.

Asif RazzaqWebsite | + postsBio ⮌

Asif RazzaqPrime Intellect Releases Prime Mega MoE: Trainable Fused MoE Kernel, 3.11x Faster Forward Pass

Asif RazzaqNVIDIA Releases Nemotron 3.5 Transcribe: An Open 30B MoE Model for 40-Language Transcription with ~4B Active Parameters

$uv pip install omni-parse-bench✓text✓tables✓layout→ scored per pageStar on GitHub→Sponsored

Baseten 发布 2B 决策模型 joey-1，每百万 token 定价 0.01 美元

推荐理由

主题

标签

正文