TL;DR: One RTX 5090, two local models, and both one-shotted a playable Three.js game. The 125B is 2× faster, but the 27B is the one I'd actually use. 🎮⚡
Same brief to both: "build a neon lane runner in Three.js."
Qwen3.8-Flash-Next 125B (IQ3_S) in Strata
• 195.8 tok/s, 22,047 tokens in 112s
• about 75% MTP draft acceptance, 128K context
• The catch is that it doesn't fit in 32GB, so it spills into system RAM and the RAM fills up
Qwen3.8-27B (Unsloth UD-Q6_K_XL) on llama.cpp b11558
• 91.16 tok/s over 50K tokens
• MTP speculative decoding with n-max 2, 74% acceptance
• 64K context, q8_0 KV cache
• about 29.5GB VRAM, fully on the GPU with no RAM spill
Flash Next is about 2.1× faster per token, but the 27B stays in VRAM and the rest of my machine stays usable. Speed matters, but so does being able to use my PC while it runs.
Rig: 5090 32GB, Ryzen 7 9850X3D, 64GB DDR5-6000, Windows 11.
llama.cpp flags for anyone who wants to try it:
-ngl 99 -fa on -c 65536 -ctk q8_0 -ctv q8_0 --jinja --temp 1.0 --top-p 0.95 --top-k 20 -md mtp-Qwen3.8-27B-Q4_0.gguf --spec-type draft-mtp --spec-draft-n-max 2 -ngld 99
Local AI on one consumer GPU is getting very real. Building games with it at Skrlr.