Jeveration int8 ONNX

ONNX export of Jeveration text-to-image for ONNX Runtime Web (WebGPU). Jeveration generates a 512 x 512 image in one forward pass of a Qwen3.5-4B backbone, so it runs in the browser without a denoising loop.

Demo: linoyts/Jeveration-WebGPU

Pipeline

  1. The prompt goes into the chat template (assets/prompt_template.json) and is embedded (embed_tokens_q4f16.onnx).
  2. 128 decision tokens are drawn from a random state (assets/) and embedded (decision_embed_fp16.onnx).
  3. One pass of the backbone over prompt + decision tokens (backbone_int8/chunk_0..7.onnx, 4 layers each). The decision tokens attend to each other in both directions.
  4. The image head turns their hidden states into a distribution over 8192 codes per position (head_fp16.onnx); all 128 codes are sampled at once.
  5. The TA-TiTok decoder turns the codes and the CLIP text features (clip_text_fp16.onnx) into the image (decoder_fp16.onnx).

Files

File Precision Size
backbone_int8/chunk_*.onnx int8 weights (MatMulNBits, block 32), LoRA merged 8 x 0.5 GB
head_fp16.onnx fp16 0.96 GB
decoder_fp16.onnx fp16 0.61 GB
embed_tokens_q4f16.onnx int4 table (from onnx-community/Qwen3.5-4B-ONNX) 0.37 GB
clip_text_fp16.onnx fp16 0.25 GB
decision_embed_fp16.onnx fp16 0.03 GB

All graphs take and return fp32. The linear-attention (Gated DeltaNet) layers use a chunked form made of matmuls, with no Scan or Loop nodes.

Accuracy

Checked against the official PyTorch pipeline (bf16, same random draws) on 4 prompts, with ONNX Runtime on CPU:

  • Full-precision export vs PyTorch fp32: relative difference 1.7e-6.
  • This int8 build: 92 to 96% of the 128 image codes identical to the official output (the bf16 reference itself agrees with fp32 on 96 to 98%).

License

Apache-2.0, as the original Jeveration weights. The backbone is derived from Qwen/Qwen3.5-4B, the decoder from TA-TiTok and the text tower from OpenAI CLIP ViT-L/14-336; each keeps its own license.

@article{liu2026jeveration,
  title   = {Jeveration: System One Models Now Generate},
  author  = {Liu, Yuhe and Li, Qi and Zhang, Shihua and Wang, Xinchao},
  journal = {arXiv preprint arXiv:2611.24738},
  year    = {2026}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for linoyts/Jeveration-int8-ONNX

Finetuned
Qwen/Qwen3.5-4B
Quantized
(1)
this model

Spaces using linoyts/Jeveration-int8-ONNX 2

Paper for linoyts/Jeveration-int8-ONNX