Kermitkiller666/Hive-9B-MLA-BF16

Hive-9B-MLA BF16

Hive is a Qwen3.5-9B vision-language base with selected language-attention layers transplanted to a custom Multi-head Latent Attention (MLA) KV design. This is the self-contained BF16 release of the trained Hive MLA checkpoint.

Contents

• model.safetensors — one 18.8 GB BF16 file containing the complete Qwen backbone, vision encoder, and trained MLA modules.
• Tokenizer, chat template, and Qwen vision preprocessor descriptors.
• qwen35_v4/ and load_hive_model.py — the required custom MLA runtime.

Loading

This is not a drop-in standard Qwen checkpoint: normal Transformers, llama.cpp, Ollama, and LM Studio loaders do not construct its MLA blocks.

pip install -r requirements.txt
python load_hive_model.py . --check-only

The model retains Qwen3.5 vision support. The runtime currently does not implement cached token-by-token generation (use_cache=False); it is a reference loader for the custom architecture, not a serving package yet.