Last updated:
Qwen4-27B: the open-weight Qwen4, before release
Qwen4-27B is the tier of Qwen4 meant for local deployment. Alibaba showed it on the Apsara keynote slide on September 22, 2026, but has not published weights, a model card, a license or a date.
What is known about Qwen4-27B
| Qwen4-27B | |
|---|---|
| Status | Announced, not released |
| Role | Open-weight tier for local deployment |
| Parameters | “27B” is the tier name; exact count and architecture not published |
| Context, modalities | Not published |
| License | Not published (Qwen3.8-27B used Apache 2.0) |
| Release date | Not announced; Alibaba said “soon” |
Source: Alibaba Cloud press release (Sep 22, 2026)
Source: DataLearner: Qwen4-27B status
Source: Hugging Face search: qwen4
The baseline: Qwen3.8-27B
The current open 27B model is the best guide to what Qwen4-27B will ask of your hardware. From its model card:
- 27B parameters, dense, with a vision encoder for image and video input.
- 64 layers in a hybrid layout: three Gated DeltaNet layers for every full-attention layer, which keeps the KV cache small.
- 262,144-token native context, extendable to about 1M with YaRN.
- Apache 2.0 license.
Source: Qwen3.8-27B model card
Hardware you will likely need
| Setup | Estimate | Fits on |
|---|---|---|
| 4-bit (Q4_K_M), 32K context | about 20 GB | 24 GB GPU, Mac with 32 GB+ |
| 3-bit (Q3_K_M), 8K context | about 15 GB | 16 GB GPU |
| 8-bit (Q8_0), 32K context | about 32 GB | 48 GB GPU, Mac with 64 GB |
| BF16 weights only | about 54 GB | 80 GB GPU |
Estimates use the Qwen3.8-27B layout. If Qwen4-27B is a plain dense model instead, the 4-bit, 32K row grows to about 26 GB because of a larger KV cache.
Work out your own numbers in the VRAM calculator
More on Qwen4
Qwen4 status and release date · VRAM calculator · GGUF downloads · Run locally · Try online · Qwen4-27B vs Qwen3.8-27B
FAQ
Is Qwen4-27B released?
No. As of October 11, 2026 there are no official Qwen4-27B weights on Hugging Face or ModelScope.
Will Qwen4-27B be open source?
It is presented as the open-weight tier of Qwen4. The license is not published yet; Qwen3.8-27B used Apache 2.0.
Can Qwen4-27B run on an RTX 4090?
Probably, at 4-bit. If it is close to Qwen3.8-27B, a Q4_K_M build with a 32K context needs about 20 GB, within a 24 GB card.
Is 27B the real parameter count?
It is the tier name. Alibaba has not published the exact parameter count or architecture.
Independent guide, not affiliated with Alibaba or the Qwen team. Status checked on October 11, 2026. Qwen4 weights are not released yet; we will update this page when they are.