Inkling 2 Mini edges Qwen 3.8 27B on coding (277 vs 272) but struggles on PC hardware

Ubunta · x · 2026-11-21

Tested on a Minisforum mini PC (Ryzen AI 9 HX 370, Radeon 890M iGPU, 89GB RAM, Linux) at 4-bit (17GB), Inkling 2 Mini shows real code quality: it beat Qwen 3.8 27B on a hard hidden-assert coding suite (277 vs 272) with cleaner code, but was slower (23 vs 16 min), used more tokens (11,704 vs 10,157), and was painful to set up — Ollama doesn't support PC yet; it needed Unsloth's llama.cpp fork on ROCm or mainline llama.cpp plus unmerged PR #32704 on Vulkan. Despite only 3.4B active params, running middle layers 3x per token caps it at 9–10.5 tok/s on a 120GB/s bus; Macs and Strix Halo hit 17–27 tok/s thanks to 2–5x bandwidth. Qwen remains the author's daily driver.

Original post →

More from Infra

Infra channel →