Paradigma

Paradigma's fellows explain their record-setting training runs: FP8 MLP down-projection with fused delayed-scale quantization, a damped late-stage weight-extrapolation optimizer, cheap attention and embedding gates, and learnable exclusive self-attention, plus lessons on steering coding agents in re

Paradigma's technical report on Limite 1B Violetto, a math-focused model trained from scratch. It covers sample-efficient architecture choices, SFT plus RL post-training with a value model, synthetic data mixes, and a released 7-benchmark eval suite.

Paradigma proposes Flywheel, a git-like graph for sharing research in progress. A Codex agent pushed a Lean-verified 2x2 result on Gröchenig's Riemann Hypothesis criterion, with caveats and failed branches published alongside, and argues review can shift toward other models.

Paradigma released Limite 1B (Violetto), a math-focused model trained from scratch on under 300B tokens. Apache 2.0 weights include the base model (59.8% few-shot MATH-500) and the value model. Aimed at single-turn, high-throughput solving in multi-agent systems; a technical report and Rainfall harn

Paradigma released Limite 1B - Violetto, a 1B-parameter math reasoning model trained on under 300B tokens that it says scores 74.25% on BeyondAIME versus 70% for the 30B MUSE-Glimmer model, released with weights, evals, and a vLLM inference plugin.