---
license: apache-2.0
base_model: Qwen/Qwen3.8-27B
tags: [taardis, ternary, base3, k-trit, w3a16, gptq, quantized, multimodal, vision]
---

# 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).