CodeMasterCody3D/taardis-27b-ubit

TAARDIS 27B ubit (~2 bpw base-3 k-trit) — vision-multimodal

Qwen3.8-27B quantized to a per-layer base-3 k-trit ("ubit") recipe — per-linear block-Hadamard rotation + Hessian-weighted GPTQ + flip-polish — placed from the k5 source, placement-only (no recon/branches/fine-tune). The vision tower + MTP head from Qwen3.8-27B are grafted at k3-g128.

Loads + numbers

• 616 packed linears + dense — 9.74 GB packed.safetensors, ~10 GB resident.
• Text PPL 6.9076 / CE 1.9326 (bundle ruler, placement-only).
• Load: ktrit_resident.load_ktrit_resident(dir, "cuda") builds the VL class (AutoModelForImageTextToText); needs the onebit-forge code (code/ktrit_resident.py + code/base3_pack.py here, plus rotation.py, ternary_gptq.py).

Vision / MTP (honest)

• Vision: live — the tower loads (model.visual) and the model instantiates as multimodal.
• MTP: weights present but inert — transformers 5.16 builds no MTP head for this arch, so the 8 mtp.* k3-packed tensors are kept in the file but not attached. MTP only accelerates speculative decoding (unused by a normal generate), so this doesn't affect chat quality; a custom MTP module (or a newer transformers) would activate them.

Caveats

Reasoning model (emits <think>); no fused high-k GEMV yet (decode-per-forward, slow long gen); a few NaN-Hessian tensors fell back to rotation+RTN; placement-only (math/logic hold, factual recall wobbles — recon/composers are the next lever). Base: Qwen/Qwen3.8-27B (Apache-2.0).