Base Labs Launches New Open-Weight Decision Models

The Mammoth-1 and Elephant-1 models aim to improve AI accuracy by utilizing a new training method for better decision-making.

Updated on Oct. 22, 2026 in Artificial Intelligence

Base Labs Launches New Open-Weight Decision Models

Live Poll

Do you prefer using open-source AI models over proprietary alternatives for your own professional projects?

Base Labs has released two new open-weight decision models, Mammoth-1 and Elephant-1, which are built on post-training Qwen3.5 architecture. These models are designed to address previous limitations in AI decision-making by prioritizing the ability to identify when a decision should be avoided.

Why it matters

Base Labs reports that earlier decision models struggled with knowing when they should not make a decision, a critical gap in reliability. By rewarding an 'I don't know' option during training, these models aim to provide more accurate and honest outcomes in complex scenarios.

The Mammoth-1 model features 27 billion parameters, while the Elephant-1 model contains 9 billion parameters. Both were trained using a mix of 6.21 million examples by altering the attention mask on full-attention layers for bidirectional visibility.

The players

Base Labs

An artificial intelligence research entity focused on developing decision-making models.

Hugging Face

A collaborative platform that serves as a central hub for hosting and sharing open-source machine learning models and datasets.

TypeSafe

The technology organization that originally introduced the decision model format using the Jev index.

The details

To achieve these capabilities, the training process incorporates a squared-error term in the loss function while randomly inserting an 'I don't know' option into the data. These models are currently available for public access through the Hugging Face platform.

TypeSafe introduced the decision model format with Jev in September 2026.
Base Labs released the mammoth-1 and elephant-1 models in October 2026.

The Tech Race

These models reflect a broader industry push to move beyond standard output and into reliable, high-stakes decision architecture. This technology replaces legacy approaches that struggled with uncertainty, effectively challenging existing standards on the Jev Decision Index.

Developers and researchers can immediately integrate these open-weight models into their workflows via Hugging Face to improve decision reliability. This release provides a new tool for those seeking models that prioritize accuracy and error identification over forced output.

The takeaway

Reliability in AI is increasingly being measured by the ability of a system to recognize when it lacks sufficient information to make a valid claim. Users should look for models that incorporate uncertainty-aware training methods to ensure safer deployment in automated processes.

What happens next

Base Labs intends to develop and release its own proprietary decision benchmark, and has indicated that multimodal decisions and looped decision models are areas currently designated for future research.