Limite 1B Runs Locally at 57 Tokens/s on M4 Pro, Scores 4/8 on AIME 2025

A mathematician ran Paradigma's Limite 1B locally on M4 Pro via an experimental MLX port at 57 tokens/s, scoring 4/8 on sampled AIME 2025; an account suggesting it is official noted AIME evals average 50k output tokens and need longer generation time.

2026-09-23 ~ 2026-09-23 · 4 related posts

• Opus 5.5 paints pixel-by-pixel in pure Python code, no image model — sergeykarayev · 2026-09-23
• No image model: Claude paints pixel-by-pixel with ~7,500 lines of Python — _sholtodouglas · 2026-09-23
• No Image Model Needed: Claude Writes Per-Pixel Python Code to Paint — repligate · 2026-09-23

1 near-duplicate retellings: alejandroll10