Robots Atlas>ROBOTS ATLAS
elephant-1
AI Models

elephant-1

1
9B decision model by Base Labs (Baseten), post-trained from Qwen3.5-9B; returns calibrated probabilities for questions over text and image inputs.
✓ Active✓ Public access⚖ Open sourceMultimodalSpecialized AI
Context window
16K
tokens
Parameters
9B
parameters
Release date
19 October 2026
Access:APIDownloadDeployment:☁ Cloud💻 Local

Overview

elephant-1 is a 9-billion-parameter decision model developed by Base Labs, the internal research team at Baseten, and post-trained from Qwen3.5-9B.

Given a state described in text (optionally with images) and a set of questions, it returns calibrated probabilities for every option in a single pass. It supports yes/no questions, single-choice selection among named options, and scoring on an ordered scale.

elephant-1 uses a block-bidirectional attention mask on its full-attention layers and has a maximum sequence length of 16,384 tokens. Weights are released under the MIT license; it has a larger sibling, mammoth-1 (27B).

Classification
MultimodalSpecialized AI
Applications
Workflow automation
Access & deployment
APIDownload
CloudLocal
Weights: Open source
Key parameters
📏 Context: 16K
🧩 Parameters: 9B
📥 Input: text, image

Technical specification

Context window
16K
tokens
Parameters
9B
parameters
License
MIT
Hardware requirements
Self-hosting requires one CUDA GPU and the pytorch/pytorch:2.8.0-cuda12.8-cudnn9-devel image; OneJev serving runtime.
Modalities
⬇ Input
textimage
⬆ Output
structured_data

Capabilities and applications

Native model capabilities
Classification
Assigning an observation to one of predefined classes (binary or multi-class). Output: class label and optionally probabilities.
Category: other
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Image understanding
Analysing and interpreting the content of images.
Category: vision
Multimodal understanding
Category: multimodal
Application domains
Workflow automation

Benchmark results

2 benchmarks
Jev Decision Index 0.4
62.65
📄 Raport techniczny Base Labs (Baseten)
Best model under 15B parameters; ranked #8 of 208 models.
Artificial Analysis Decision Index (replika Base Labs)
65.2
📄 Raport techniczny Base Labs (Baseten)
Level with the d1-ultra model (65).

Technical architecture

Core Architecture
TRTransformer
Model Form
MLMultimodal LLM
Training Techniques
POPost-TrainingFIFine-tuning

Sources and related pages