Intel code leaks Qwen 4 for 1 graphics card

Code for Intel graphics cards named an unreleased Alibaba model, Qwen4-27B, ready for 1 card, IAMag reported on the 24th of October 2026. Alibaba has not confirmed the details.

The code change, called a pull request, was merged on the 23rd of October. It runs the model on 1 or 2 cards in a squeezed 4-bit form, and on 2 cards in a larger form.

27B means 27 billion parameters, the numbers a model learns in training. The code calls it a dense model, meaning all its parts work on every word.

What the code says

Tests ran on 8 Intel B70 cards on the 22nd of October, IAMag reported. The author wrote that "Qwen4 QSA weights were not provided", so the real model was never tested.

A mystery model on Arena

A model called Kalmia appeared on Arena, a public site where people compare AI answers, on the 21st of October. Testers believe it is Qwen 4, but Alibaba has not said so.

In one tester's Minecraft building test, Kalmia placed 23,750 blocks and Claude Sonnet 5.5 placed 44,107. The tester, jan_volad, wrote that Sonnet "did way better".

Below we break down the numbers

The code's early numbers show the wait for the first word falling from about 9.1 to 5.2 seconds. That is about 43 percent faster, by our arithmetic.

What is confirmed and what is not

On the 27th of October the code page no longer opened, and GitHub showed the whole project as not found. The details above come from IAMag and Huxiu, which read the page earlier.

A llama.cpp request called "Get ready for qwen4", opened on the 24th of October, still links to it. Alibaba said in September that Qwen4-27B is coming, IAMag reported, with no date.

Key facts

Intel graphics card code named an unreleased Qwen4-27B model.

The code sets it up for 1 or 2 cards in a squeezed form.

The real model was never tested in that code.

Alibaba has given no release date.