limite-1b-violetto-mlx

Limite 1B - Violetto (MLX)

This repository contains the Apple Silicon (MLX) formatted weights and architecture implementation for Limite 1B - Violetto, created by Paradigma.

Full source code, tests, and documentation: github.com/gabe-rbo/mlx-limite-1b-violetto

Overview

Limite 1B - Violetto is an open-source 1-billion parameter dense autoregressive transformer designed for high-throughput, competition-level mathematical reasoning.

• AIME 2026: 94.01%
• BeyondAIME: 74.25%
• HMMT Feb. 2026: 83.62%
• AIME 2025: 90.21%

Running with mlx-lm

Install mlx-lm:

pip install mlx-lm

Run in Python:

import mlx_lm

model, tokenizer = mlx_lm.load("gabe-rbo/limite-1b-violetto-mlx")

prompt = (
    "<|im_start|>system\n"
    "You are a helpful assistant.\n"
    "Please reason step by step, and put your final answer within \\boxed{}.<|im_end|>\n"
    "<|im_start|>user\n"
    "Solve for x: x^2 - 5x + 6 = 0.<|im_end|>\n"
    "<|im_start|>assistant\n"
)

response = mlx_lm.generate(
    model=model,
    tokenizer=tokenizer,
    prompt=prompt,
    max_tokens=2048,
    temp=0.6,
    top_p=0.95,
)
print(response)

Running as a Local Server (LM Studio / Bionic / OpenAI API compatible)

python -m mlx_lm server --model gabe-rbo/limite-1b-violetto-mlx --port 8080 --temp 0.6 --top-p 0.95

Point LM Studio, Bionic, or any OpenAI-compatible app to http://localhost:8080/v1.

Citation

@misc{paradigma2026limite,
  title = {{Limite 1B - Violetto}},
  author = {Prignano, Mario and Cirillo, Gabriele and
            Morosini, Alessio and Cerovaz, Luca and
            Bartolocci, Alessandro and Rodolà, Emanuele and
            Starace, Giulio and Pappone, Francesco},
  year = {2026},
  howpublished = {\url{https://paradigma.inc/blog/limite-1b-violetto/}}
}