Last updated:
Qwen4: what is confirmed and what is still rumor
At its Apsara Conference on September 22, 2026, Alibaba said Qwen4 "is currently in training." Keynote coverage named four tiers: Qwen4-Max, Qwen4-Plus, Qwen4-Flash and Qwen4-27B, the open-weight tier for local use. As of October 11, 2026 there are no Qwen4 weights, model card, license, price or release date.
What is confirmed so far
| Item | Status on Oct 11, 2026 |
|---|---|
| Qwen4 is in training | Confirmed in Alibaba’s Apsara press release (Sep 22, 2026) |
| Four tiers: Max, Plus, Flash, 27B | On the keynote slide, reported by attendees; not in the written press release |
| Qwen4-27B open weights | Expected: the 27B is presented as the open-weight tier. No weights yet |
| Qwen4-Max open weights | Not confirmed |
| Release date | Not announced. Alibaba said “soon” |
| Parameters, context, license, price | Not published for any tier |
| Qwen4 on Hugging Face | No Qwen4 repository under the Qwen organization |
Source: Alibaba Cloud press release (Sep 22, 2026)
Source: DataLearner: Qwen4-27B status and keynote sources
Source: Hugging Face search: qwen4
About “Qwen4-27B and Qwen4-Max will be open-sourced together soon”
Posts on X say Alibaba teased Qwen4-27B and Qwen4-Max for an open release together, soon. We could not find an official Qwen or Alibaba post with that wording or a date. The official record says Qwen4 is in training, and the only tier described as open-weight is the 27B.
Treat any specific date as a guess until weights appear on Hugging Face or the Qwen blog. Launch dates for new models often slip.
The four Qwen4 tiers
| Tier | Role | Open weights? |
|---|---|---|
| Qwen4-Max | Flagship | Not confirmed |
| Qwen4-Plus | Balanced | Not confirmed |
| Qwen4-Flash | Fast, high volume | Not confirmed |
| Qwen4-27B | Local deployment | Expected |
Roles come from keynote coverage. Alibaba has not published specs for any tier.
Source: DataLearner: Qwen4-27B status and keynote sources
What Qwen4’s architecture probably looks like
In August 2026 Qwen released Qwen3.8-Flash-Next. Its model card calls it “an experimental preview of the architecture that will underpin Qwen4.”
- Mixture of experts: 125B total parameters with about 6B active per token, plus 51B N-gram embedding parameters.
- Hybrid attention: Gated DeltaNet combined with Qwen Sparse Attention (QSA).
- 262,144-token native context, extendable to about 1M with YaRN.
- License: qwen-community-1.0, not Apache 2.0.
This is a preview model, not Qwen4. Nothing official says Qwen4-27B will share these numbers.
Source: Qwen3.8-Flash-Next model card
Signs the release is getting close
A vLLM pull request titled “[Kernel][Qwen] Add qwen4 fuse op” adds kernels for this architecture. Inference engines usually add support shortly before or at launch, so this points to a release being prepared, but it does not give a date.
Source: vLLM pull request: Add qwen4 fuse op
How to get ready for Qwen4-27B
- Check your memory. If Qwen4-27B is similar to Qwen3.8-27B, a 4-bit build with a 32K context needs about 20 GB, so a 24 GB GPU or a Mac with 32 GB or more is the comfortable minimum. Full BF16 weights alone are about 54 GB.
- Install a runtime now: Ollama or LM Studio for the easiest start, llama.cpp for control, vLLM or SGLang for serving.
- Run Qwen3.8-27B as a stand-in to test speed and prompts, then switch the model name when Qwen4-27B ships.
- Download GGUF files only from the Qwen organization or well-known quantizers, and check the base model on the model card.
VRAM calculator · Qwen4-27B · Qwen4 GGUF downloads · Run Qwen4 locally · Try Qwen4 online · Qwen4-27B vs Qwen3.8-27B
FAQ
When is Qwen4 coming out?
There is no date. On September 22, 2026 Alibaba said Qwen4 is in training and coming soon. As of October 11, 2026 no Qwen4 model has been released.
Will Qwen4 be open source?
The 27B tier is presented as the open-weight tier for local use. Its license is not published. Open weights for Max, Plus and Flash are not confirmed.
Can I download Qwen4 now?
No. There are no official Qwen4 weights yet. Files labeled “Qwen4” that appear before the official release are not from Alibaba.
How much VRAM will Qwen4-27B need?
Unknown until release. If it is similar to Qwen3.8-27B, a 4-bit build with a 32K context needs about 20 GB, and the full BF16 weights are about 54 GB.
Is Qwen3.8-Flash-Next the same as Qwen4?
No. It is an open preview of the architecture that Qwen4 will build on, released in August 2026.
Independent guide, not affiliated with Alibaba or the Qwen team. Status checked on October 11, 2026. We will update this page when Qwen4 weights are released.