florianvoss/Qwen3.5-4B-INT8-Modalix

Qwen3.5-4B INT8 — LLiMa runtime package

Compiled model artifacts for the LLiMa runtime on SiMa.ai Modalix.

• BF16 activations and INT8 weights (A_BF16_W_INT8).
• Maximum token count: 4096; prefill group size: 128.
• Compiled vision input: 32 × 32.
• Filter sharing, quantized embeddings, and quantized KV cache enabled.

Run

Copy or download the entire repository, preserving the directory structure. On a Modalix device with a compatible LLiMa runtime installed:

llima run /path/to/Qwen3.5-4B-INT8

devkit/ contains model configuration, tokenizer files, and embeddings. elf_files/ contains the compiled MLA programs. These are device runtime artifacts; they are not a Transformers checkpoint.

Compilation and local deployment completed. This package has not been validated by running the full model on a Modalix device.