SYLIN
Language & code models   WK 47 / 2026   Updated Nov 19   1 min read

Same GPU, same models, up to twice the tokens per second

ComfyUI 0.45 and llama.cpp 1.0 speed up local LLMs, though ComfyUI's numbers are its own.

2X FASTER

Update ComfyUI to 0.45 and the same Qwen3.5 9B checkpoint went from 106 to 219 tokens per second on an RTX 5090. A tokens-per-second figure is how fast the model writes. This week's local stack also brought llama.cpp 1.0 and fresh Unsloth quants.

The short version
Good atfaster generation from models you already have, and running MoE models larger than your VRAM.
To try itall free. ComfyUI 0.45, llama.cpp 1.0, Unsloth's Qwen3.8-27B GGUFs, Hugging Face's new Llama desktop app.
The catchthe speedups are vendor figures. The SSD-streaming report is one user's claim.

The deeper look

ComfyUI rewrote its text-generation path for Qwen3.x. Its blog reports Qwen3.5 9B going from 106 to 219 tok/s on an RTX 5090 and from 91 to 129 on an RTX 5080, same quality. Nobody independent has reproduced that yet.

llama.cpp 1.0 declares its API stable, removes 35 deprecated functions, and adds a hot-expert cache. A mixture-of-experts (MoE) model only uses a few of its experts per token. The cache keeps the busy ones in VRAM, so models bigger than your card get less painful.

Unsloth Dynamic GGUF quants v3.1 GGUFs for Qwen3.8-27B claim 5 to 10% better accuracy than rival quant methods. Unsloth itself notes that rivals dispute such comparisons. TokenWing streams experts from SSD, and one Reddit user reports 45 to 50 tok/s on a 120B MoE with 12 GB VRAM and 32 GB RAM. One user, one machine.

It is for tinkerers who already run models locally. Wait for independent numbers before restructuring your setup.

How to start: update ComfyUI to 0.45, or grab llama.cpp v1.0.0 from its GitHub releases. For a no-terminal route, try the Llama app at llama.app.

How to start
From the code
$ git clone https://github.com/comfyanonymous/ComfyUI

More this week5 stories