Ornith 1.5 122B runs on one 32 GB card, but the 35B is within noise of it on every test…
This is a dev post classified by AGK Intelligence as Other (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
Ornith 1.5 122B runs on one 32 GB card, but the 35B is within noise of it on every test except MMLU, and 6x faster. TL;DR: - 122B, 3-bit file (46.1 GiB): 47.6 tok/s on one RTX 5090, with part of its experts in system RAM - 35B, Q4_K_M (20.2 GiB): fits fully on the card, 303 tok/s - first token on a 16k prompt: ~28 s vs ~2 s - GPQA Diamond, thinking on: 83.3 vs 81.8, a tie on 198 questions - MMLU is the one clear 122B win: 85.8 vs 82.2 take: on a 32 GB card, get the 35B. the 122B is worth the wait only if general knowledge matters more to you than speed. full report, with the offload sweep a
Posted by witcheer (22.8k followers) 3 days ago · 50 likes · 3.7k views · view the original post on X. Kept by the Dev Radar as Other. Tools mentioned: llm-bench-rig.
More dev work like this
- Just added my signature on the Denim upgrade for the @base security council. — @toady_hawk
- 强化学习做交易，很多人入门撞上的第一个框架就是FinRL，你大概也不例外。 — @bkdgiffug
- Deepseek v4.1 Flash 1017 hitting as high as 72 tok/s decode on repetitive tokens — @Italianclownz
- With Antioch, you can simulate any deployment environment for your robot — even the… — @antiochrobotics
- took @rehan_shei's open-source pixel fx into a real 3D scene and made the wizard… — @majidmanzarpour
- 彭博终端贵得让个人用户直摇头，这感受你应该不陌生。 — @bkdgiffug
- I went quiet for a few weeks because what I’m building now is better than every library… — @dwhitedesign
- Celestia’s Voyager handled 21M transactions per second in its first big test — @BSCNews
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.2k posts from 5k X accounts over the last 21 days, 8.8k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-10-30 22:12 UTC. Full method.