Ornith-1.5-122B MoE runs on a DGX Spark: 52 tok/s, 256K context recipe
lmoroney · x · 2026-10-22
Ornith released Ornith-1.5-122B, a mixture-of-experts model built for coding and agent tasks, reportedly adding pre/mid/post-training on top of Qwen3.5. SGLang shipped day-zero support with reasoning, tool calling, and 262K context configs.
@MiaAIlab published a full recipe to run it on a DGX Spark using a 4 bpw EXL3-EAQ quant (adaptable to RTX PRO 6000):
- 256K context, 1M total KV cache, 4 concurrent streams by default
- 86 GB total memory, DFlash2 drafts
- Decode (thinking off): 52 tok/s prose single-stream, 103 tok/s code, 90 tok/s prose at 4 streams
- Prefill up to 2000 tok/s
The author notes it "feels different from Qwen3.8-Flash-Next and does things its own way" — a fresh data point for how big an open model can get on a desk.
More from Infra
- BIND9 hit by high-severity GSS-TSIG replay bug enabling potential RCE (CVSS 8.1) — basedjensen · 2026-10-22
- Who priority-parks background agents? An ops veteran on sharing one 16GB GPU between apps and agents — Powermadlux · 2026-10-22
- Gartner: AI data centers need $8.5T revenue by 2030 but will only make $3.6T — AIFlow_ML · 2026-10-22
- FlashAttention 4 on GB300 Is So Fast It Annoys Performance Researchers — francoisfleuret · 2026-10-22
- Comment: Larry Page worth $300B won't pay 5% of net worth for infrastructure he profited from — georgemillo · 2026-10-22
- Amazon reportedly buying $5B of Nvidia GB300 racks it can't yet power on — AIFlow_ML · 2026-10-22