Experimental model. This repository is an experimental Alkahest/Rally package. It may fail, behave unpredictably, or produce unsuitable output. Use at your own risk; do not rely on it for safety-critical or production decisions without your own validation.

Alkahest 0.8B Q4 ONNX

Public experimental browser-oriented ONNX package for the finalized Alkahest 0.8B direct lane.

Status

  • Variant: direct Alkahest 0.8B
  • Runtime: Transformers.js / WebGPU
  • Preferred dtype map: { "embed_tokens": "q4", "decoder_model_merged": "q4", "vision_encoder": "fp16" }
  • Browser smoke: passed text load, WebGPU session initialization, and coherent text generation

Files

  • onnx/embed_tokens_q4.onnx
  • onnx/embed_tokens_q4.onnx_data
  • onnx/decoder_model_merged_q4.onnx
  • onnx/decoder_model_merged_q4.onnx_data
  • onnx/vision_encoder_fp16.onnx
  • onnx/vision_encoder_fp16.onnx_data

Notes

This package includes Qwen3.5 RMSNorm offset patching during ONNX weight transplant. Older experiment repos without that patch can load but generate corrupted text.

Downloads last month
4
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support