Language & code models   WK 47 / 2026   Updated Nov 19   2 min read

Your local 27B model just got faster kernels, better quants and a coding adapter

Qwen3.8-27B now has a full local stack around it, though most of the numbers are the makers' own.

27B AT HOME

Qwen 3.8 at 27B is turning into the model people run at home, and this week the tools around it all moved. ComfyUI 0.45 reports up to double the generation speed, Unsloth shipped new quants, and two adapters aim it at coding agents.

The short version
Good ateveryday chat and coding on one consumer GPU. The weights are Apache 2.0.
To try itfree. Unsloth's Dynamic v3.1 GGUFs run 2- to 6-bit, and one user runs Q4 on a $1,500 box with 32 GB of VRAM.
The catchnearly every speed and accuracy figure comes from the maker of the tool. The Pi adapter's results were already corrected once.

The deeper look

Comfy Org's blog reports its 0.45 rewrite doubling tokens per second: 219 against 106 for the 9B model on an RTX 5090, and Qwen3.8 27B describing an image going from 55 to 122. That is ComfyUI measuring itself. Unsloth Dynamic GGUF quants claims its v3.1 GGUFs score 5 to 10% higher in accuracy than rival quantization methods, and notes that rivals dispute such comparisons.

The adapters are the new part. Hugging Face's OpenCode adapter lifts the hard SWE-bench Pro subset from 35.3% to 51.0% with thinking off. The Pi adapter first claimed 15 to 32 solves out of 51 hard tasks, then corrected it to 20 to 33 after a tunnel outage hit nine base runs. Both are Hugging Face's numbers, so expect smaller gains on your own repos. On a single H100, Tencent's DSpine reached about 150 tok/s against 129 for DFlash 2, per a hands-on test by Fahd Mirza.

The human evidence is one comment: a user says Qwen3.8-27B at Q4 replaced about 95% of their conversational Claude use. One user, one workload. A Qwen4 27B is reportedly coming soon.

This is for tinkerers who already run models locally. Skip it if your card has under 16 GB or you want a hosted frontier model.

How to start: pull the Dynamic v3.1 GGUFs from unsloth.ai/docs/models/qwen3.8, update ComfyUI to 0.45, and for coding try harness-hub/Qwen3.8-27B-opencode-adapted on Hugging Face.

More this week5 stories