Radar· local AI & agent security
models low confidence

Qwen4: no weights yet, but signs of a 27B (and maybe a 35B) are getting stronger

Nov 18, 2026 · qwen4, rumor

TL;DR Qwen4-Max is still API-only. Huawei published Qwen4-27B quants and a Qwen team lead hinted at smaller models.

In plain words: everyone running AI at home is waiting for a mid-size Qwen4 that fits on one or two graphics cards. It is not out yet; what we have are hints. Until it appears under the official account, treat any “Qwen4” download as suspicious.

Where things stand (Nov 18)

The Qwen organization on Hugging Face has no Qwen4 weights yet. Qwen4-Max is available only through the API, and the earlier preview is retired on Nov 25.

The signals

  • Huawei added Qwen4-27B quants to its MACaQ repository, with support in mainline llama.cpp. Estimated sizes: ~6.75 GB at 2-bit, ~10.1 GB at 3-bit, ~13.5 GB at 4-bit.
  • A Qwen team lead replied with a line read by the community as a hint at a possible Qwen4-35B.
  • Unsloth says its Qwen4 GGUFs will ship with Dynamic 3.1, tuned for coding agents.

Why it matters

A 27B at ~13.5 GB in 4-bit fits a single 16–24 GB GPU with room for context.

What to do

This is a rumor until it shows up under Qwen/. Third-party “Qwen4” repositories appeared before the launch and were fake. Download only from the official organization and check the hash.

Sources

← back