ultraviolet-50m-mid-alpha

Ultraviolet 50M (UV-50M) is a 50.8 million parameter decoder-only transformer model trained primarily on math for rapid RL experimentation on language models. It is intended to be a small model organism for reinforcement learning which can be effectively trained on any consumer GPU.

UV-50M's name and architecture are derived from the Limite 1B Violetto model produced by Paradigma. UV-50M is optimized for inference speed rather than maximal performance per parameter. This allows reinforcement learning experiments to run at high speed on consumer GPUs: at short context lengths (under 1k tokens) it exceeds 200k tokens per second of throughput on a single 2080 Ti.

This checkpoint is the math reasoning model: ultraviolet-50m-base-alpha fine-tuned on <think> reasoning traces (violetto-math-traces, config mid-alpha). It thinks, closes </think>, then answers in \boxed{}.

parameters	50.8M (10 layers, d_model 640, 8 query heads / 1 KV head, head_dim 64)
context	4096 tokens
tokenizer	16,384-token BPE, in tokenizer/
training data	sources and decontamination in DATA.md
format	model.pt, a torch checkpoint dict (model = fp32 state dict, cfg = model config)
MATH-500	pass@1 0.090, pass@16 0.356 (T = 1.0, 16 samples, max 2048 new tokens)
GSM8K test	pass@1 0.067, pass@16 0.390

Usage

Model code: github.com/hliuson/ultraviolet.

import torch
from huggingface_hub import hf_hub_download
from ultraviolet.model import MathLM, MathConfig

s = torch.load(hf_hub_download("harribo11/ultraviolet-50m-mid-alpha", "model.pt"), map_location="cpu", weights_only=False)
model = MathLM(MathConfig(**s["cfg"]))
model.load_state_dict(s["model"])
tokenizer = hf_hub_download("harribo11/ultraviolet-50m-mid-alpha", "tokenizer/tokenizer.json")   # load with `tokenizers`

Prompt: <|endoftext|>Problem:\n{problem}\n\n<think>\n. The model writes its reasoning, </think>, then a solution ending in \boxed{answer}.

Citation

@misc{ultraviolet50m,
  title  = {ultraviolet: a 50M-parameter math language model},
  author = {hliuson},
  year   = {2026},
  url    = {https://github.com/hliuson/ultraviolet}
}