Gamut is a lightweight, sub‑90M‑parameter model that evaluates and optimizes design outputs from large language models. It can serve as a reward signal in reinforcement‑learning pipelines, rank multiple LLM samples, or act as an objective function to directly improve designs within a parametrized configuration space, delivering near‑ceiling performance on diverse image‑pair benchmarks.
Funding
Funding not disclosed
Founders
Product
Problem
Creating reliable code and design artifacts from markdown specifications often requires large, resource‑intensive language models, making it difficult to integrate such capabilities into constrained environments or specialized workflows. Additionally, evaluating and ranking multiple LLM outputs or using LLMs as objective functions in design optimization can be inefficient without a lightweight, high‑accuracy model.
Solution
Gamut offers a compact 90‑million‑parameter language model that translates markdown specifications into high‑quality TypeScript APIs, C libraries, load balancer configurations, and similar artifacts with near‑ceiling accuracy on a diverse 160‑pair benchmark. Its efficiency enables deployment in settings where compute resources are limited, while its strong performance makes it suitable as a reward signal in reinforcement‑learning pipelines, a judge for selecting the best among multiple LLM samples, or an objective function for optimizing parametrized design spaces. Users can evaluate Gamut on their own image‑pair datasets to verify real‑world effectiveness before integration.
Target Audience
Primary users are software developers, ML engineers, and product teams that need efficient, high‑accuracy code generation from specifications, as well as researchers building RL pipelines or design‑optimization systems requiring lightweight LLM reward or objective functions.
Features
- 90M‑parameter model optimized for markdown‑to‑code translation across multiple programming languages and infrastructure configurations
- Near‑ceiling benchmark performance (95 % confidence interval covering 154–160 of 160 test pairs)
- Low computational footprint suitable for on‑device inference or integration into constrained pipelines
- Can serve as a reward model in reinforcement‑learning workflows to guide policy improvement
- Functions as an automated judge to rank and select the highest‑quality outputs from competing LLMs
- Acts as an objective function for direct optimization of parametrized design spaces
- Provides an evaluation interface allowing users to test the model on custom image‑pair datasets