P-JiT 21K, compiled for the browser

Linum's P-JiT 21K and the Qwen3.5-4B layers it reads captions with, compiled into two fusor-ai bundles: the weights the browser runtime of pjit-21k-wasm loads to turn a caption into an image on the device's CPU. Try it in the Space fusorlabs/pjit-21k-wasm.

  • Sources: Linum-AI/pjit-21k at revision 251725ad9a0618f0a4ba3cdf4a488d35312be352 and Qwen/Qwen3.5-4B at revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a, compiled with the mix profile into two bundles, because one WebAssembly memory holds at most 4 GiB:
    • text/: the Qwen3.5-4B tokenizer, embeddings and layers 0 to 27. The embeddings and the feed-forwards' gate and up projections are 4-bit, every other matrix 7-bit (3266 MB).
    • image/: the P-JiT 21K diffusion transformer, every large matrix 7-bit (2626 MB). The patch embedding, the timestep MLP, the text projection and the head stay in f32.
  • Format: each manifest.json lists every tensor and part; each part is named by its BLAKE3 hash, which the runtime checks as it loads.
  • Reproduce it: just model in pjit-21k-wasm downloads the checkpoints and compiles the same bundles.

This is not a PyTorch or safetensors checkpoint: it only loads in the pjit-21k-wasm runtime.

Both models are Apache-2.0; Qwen3.5-4B's license is LICENSE.

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. πŸ™‹ Ask for provider support

Model tree for fusorlabs/pjit-21k-web

Finetuned
(4)
this model

Space using fusorlabs/pjit-21k-web 1