---

license: mit

base_model: ornith-ai/Ornith-1.5-35B-A3B

tags: [gguf, pxa, pxq, quantized, tensor-split, p100, v100, pascal, volta]

---

Ornith-1.5-35B-A3B — PXQ4 (tensor-split ready)

A PXQ4 quantization of ornith-ai/Ornith-1.5-35B-A3B for

the PXA engine on Tesla P100 / V100 / GTX 1080 Ti.

Works with the tensor split (-sm tensor). Every weight is PXQ4 except ssm_out, which stays at Q8_0: the

tensor split has to cut ssm_out along a dimension PXQ4 panels cannot be cut on, so a file with a PXQ4 ssm_out

is refused under -sm tensor (it still runs under -sm layer).

| | |

|---|---|

| File | Ornith-1.5-35B-A3B-PXQ4.gguf, 19,587,444,992 bytes |

| SHA256 | d5d78d3ee7cce45ef88aa999796fdc6501e1bdf5a7ab55f2aedd62664af2cd1c |

| Engine | PXA v2026.09.20 (commit f2f1d965) or later |

Run it on two identical cards (tensor split)

PXA_TSPLIT_UNPROVEN_ARCH=1 PXA_TSPLIT_REDUCE=fused PXA_TSPLIT_REDUCE_PREFILL=1 \
llama-server -m Ornith-1.5-35B-A3B-PXQ4.gguf -ngl 99 -c 262144 -np 1 -fa on \
  -sm tensor -ts 1,1 -b 2048 -ub 2048 --jinja

PXA_TSPLIT_UNPROVEN_ARCH=1 is needed because this architecture is not yet on the engine's measured list. This is

the exact configuration of our own seats: two Tesla P100 16 GB cards, full 262k context. Or let the launcher decide:

./pxa-launch --model Ornith-1.5-35B-A3B-PXQ4.gguf --gpus 0,1.

Community: PXA Network Discord

Expert-cache routing counts for PXA v2026.10 (lets the model run when it does not fit your cards); keep it next to the GGUF.