---

license: apache-2.0

base_model:

• Zyphra/ZAYA1-base

tags:

• gguf
• zaya
• moe
• base-model

---

ZAYA1-base — GGUF

Our own GGUF conversion of Zyphra's ZAYA1-base. It keeps Zyphra's original Megatron-style checkpoint layout, which transformers cannot load; our converter maps it to ZAYA's standard layout (llama.cpp #14). Converting Zyphra/ZAYA1-8B-legacy this way gives tensors byte-identical to the ones from Zyphra/ZAYA1-8B.

Contents

• ZAYA1-base-Q4_K_M.gguf (8B shape: 40 layers, 16 experts; rope theta 1e6; 32,768-token context).

Validation

• Wikitext-2 test perplexity, 60 chunks of 512 tokens, Q4_K_M on Vulkan (Radeon 8060S): 8.53 ± 0.17.
• Raw-text perplexity favours base models; the post-trained ZAYA1-8B scores 32.13 on the same test.

Running it

With the 1bit engine:

1bit serve -m ZAYA1-base-Q4_K_M.gguf --device vulkan

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-base, 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.