Base Labs Releases New Open-Weight Decision Models
The mammoth-1 and elephant-1 models are trained to avoid bad calls by learning when to opt for an I don't know response.
Updated on Oct. 22, 2026 in Artificial Intelligence
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, the 27-billion-parameter mammoth-1 and the 9-billion-parameter elephant-1. Built on Qwen3.5, these models are designed to improve decision-making accuracy by training the AI to recognize when it lacks sufficient information.
Why it matters
Earlier decision-making models frequently struggled with identifying situations where they should abstain from making a choice. By training models to opt for an 'I don't know' response, Base Labs aims to reduce overconfident or incorrect decision outputs.
Mammoth-1 features 27 billion parameters, while elephant-1 uses 9 billion. Both were trained on a mix of 6.21 million examples using a specialized loss function.
The players
Base Labs
An AI research lab focused on developing decision-making models and open-weight architectures.
Qwen3.5
A series of large language models used as the foundational architecture for the mammoth-1 and elephant-1 decision models.
The details
The models operate by modifying the attention mask on full-attention layers to allow bidirectional visibility between state information and questions. During training, researchers injected an 'I don't know' option into the data and applied a squared-error term in the loss function to penalize incorrect guesses. This process incentivizes the model to effectively communicate uncertainty rather than attempting to force a decision.
Timeline
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 enter a field currently occupied by 208 other entries on the Jev Decision Index. Base Labs is positioning these releases as a solution to the limitations of previous models that lacked explicit training to handle ambiguous prompts.
Developers can access the mammoth-1 and elephant-1 models now on Hugging Face for implementation in decision-support workflows. Users should note that these models are best suited for tasks where identifying lack of information is critical to output safety.
The takeaway
These releases suggest a shift toward training models to explicitly signal ignorance, a critical step for high-stakes AI deployment. Developers should track the forthcoming Base Labs decision benchmark to see how these models perform against future iterations of decision-focused architectures.
What happens next
Base Labs has indicated plans to release its own independent decision benchmark in the future.
Further reading
For more on the current state of large language model capabilities, visit Artificial Intelligence.
Source note: This article includes information reported by Startup Fortune.
Live Poll
Do you prefer using open-source AI models over proprietary alternatives for your own professional projects?
