Tkaa/Merkaba01-Gemma4-12B

Merkaba01 — Gemma 4 12B Runtime Core Package (v0.1 Alpha)

This repository contains the pre-packaged Merkaba01 runtime weights for Gemma 4 12B (HybridQ8Q4 .qcore + 1.58-bit .merkaba with Skeleton-DoRA).

Note: These files use Merkaba01's custom binary container formats (.qcore, .merkaba, .mbin) and are designed exclusively for the Merkaba01 Pure-Rust Cognitive Runtime Engine. They cannot be loaded directly in transformers, vLLM, or llama.cpp.

Included Files

• gemma4_12b_hybrid_q8q4.qcore (6.80 GiB) — Zero-copy mmap quantized core (Q8 + Q4_G32 + Anchor-DoRA).
• gemma4_12b_skel_dora.merkaba (222.9 MiB) — 1.58-bit ternary MLP & shared SVD anchor package (SharedAnchorsSvd + Skeleton-DoRA).
• model.safetensors (1.88 GiB) — Base runtime vocabulary embeddings.
• model_norms.safetensors (1.54 MiB) — Layer normalization and shape-stub skeleton.
• tokenizer.json & tokenizer.mbin — Fast pre-compiled binary vocabulary & SentencePiece tokenizer.

License & Attribution

Base model weights are derived from Google Gemma 4 12B and distributed under the Apache 2.0 License.