MIXAR: Scaling Autoregressive Pixel-based Language Models to Multiple Languages and Scripts
Paper • 2604.11575 • Published
MIXAR
Multilingual pixel-based language models from the MIXAR paper, plus the MGPT-2 baseline. Code: github.com/april-tools/mixar
Note The MIXAR paper. Code: github.com/april-tools/mixar
TYTTYTTYT/llama-pixar-477m-p32-stage2
Text Generation • 0.5B • Updated • 223Note Recommended. 477M, 32x32 patches, stage 2 (GRPO). The strongest generator.
TYTTYTTYT/llama-pixar-116m-p32-stage2
Text Generation • 0.1B • Updated • 211Note 116M, 32x32 patches, stage 2 (GRPO).
TYTTYTTYT/llama-pixar-116m-p8-stage2
Text Generation • 0.1B • Updated • 209Note 113M, 8x8 patches, stage 2 (GRPO).
TYTTYTTYT/llama-pixar-477m-p32
Text Generation • 0.5B • Updated • 216Note 477M, 32x32 patches, stage-1 backbone.
TYTTYTTYT/llama-pixar-116m-p32
Text Generation • 0.1B • Updated • 208Note 116M, 32x32 patches, stage-1 backbone.
TYTTYTTYT/llama-pixar-116m-p8
Text Generation • 0.1B • Updated • 203Note 113M, 8x8 patches, stage-1 backbone.
TYTTYTTYT/llama-pixar-116m-p8-stage2-gan
Text Generation • 0.1B • Updated • 211Note Ablation: adversarial (GAN) stage 2 of the 113M 8x8 model.
TYTTYTTYT/mgpt2-124m-mc4
Text Generation • 0.1B • Updated • 253Note Baseline: GPT-2 124M pretrained on the same mC4 corpus.