---

license: apache-2.0

base_model:

• Zyphra/ZAYA1-74B-preview

tags:

• gguf
• zaya
• moe

---

ZAYA1-74B-preview — GGUF

Our own GGUF conversion of Zyphra's ZAYA1-74B-preview: 74.8B parameters, 60 layers alternating sliding-window attention (a 4,097-token window, rope theta 1e4) and full attention (rope theta 1e7). Sliding-window support was added in our llama.cpp fork (PR #11).

Contents

• ZAYA1-74B-preview-Q4_K_M.gguf (45.7 GB), converted with --remote to Q8_0 and requantized to Q4_K_M. Includes the model's chat template.

Validation

• The engine's end-to-end serve test on Strix Halo (--device vulkan --ctx-size 8192): pass. Chat with the model's own template gives coherent answers (code, prose, a one-line fact).
• Measured with llama-bench on Vulkan (Radeon 8060S): prefill 505.8 tok/s (pp512), decode 35.4 tok/s (tg128).
• There is no reference check against transformers: a 74B FP32 run needs about 300 GB of memory.

Running it

With the 1bit engine:

1bit serve -m ZAYA1-74B-preview-Q4_K_M.gguf --device vulkan --ctx-size 8192

It needs about 46 GB of GPU-visible memory.

ZAYA runs from our llama.cpp fork (branch 1bit/hrx-vulkan-patched); upstream llama.cpp has no ZAYA

model. The GGUFs must come from our converter: it writes the grouped convolution's weights

tap-major, which the graph expects.

Attribution

• Base model: Zyphra/ZAYA1-74B-preview, Apache 2.0.
• GGUF conversion and validation: the 1bit engine project, with the converter and model code in our llama.cpp fork.
• License: Apache 2.0, inherited from the base model.