Qwen3.8-27B Release Date: Everything We Know So Far

Alibaba promised Qwen3.8-27B's open weights for the week of August 10, 2026, and that window came and went. Now the delay has a concrete target: the official Qwe​n page on ModelScope is running a countdown to midnight on August 15, 2026 — 00:00 JST, per PC Watch's August 13 report of the teaser. The weights still have not appeared on the official Qwe​n organization on Hugging Face as of today, but the release-date question now has an answer for the first time since the August 3 announcement. Here is what is confirmed, what is still unverified, and what to watch.

What is confirmed

Alibaba announced the Qwe​n3.8 generation on August 3, 2026. Two things happened that day: Qwen3.8-Max, the 2.4-trillion-parameter mixture-of-experts flagship, went live with a priced API, and Alibaba committed to open weights for both the Max and the smaller Qwen3.8-27B "within about a week." Multiple independent outlets (36Kr, TrendForce, C114) carried the same commitment: Hugging Face and ModelScope, the week of August 10.

The 27B is the self-hostable member of the generation — Alibaba positioned it as sized for single-GPU and on-premise work, the direct successor to the widely used Qwen3.6-27B. That lineage matters. Qwen3.6-27B became one of the most popular local coding models of its generation, and the 3.8 version inherits that installed base of questions: what does it cost to run, is it better at agents, when can I pull it.

What is independently verifiable today is the flagship, and that now includes open weights. Qwen3.8-Max is live on API, and its repositories — Qwen3.8-2.4T-A95B (full BF16) and an official FP8 quantization — went up on the official Qwe​n organization on Hugging Face in the week of August 10, the first time Alibaba has open-sourced a Max-class flagship. It is already scored too: Artificial Analysis puts its Intelligence Index at 58, the closest a Chinese lab has come to the top of that table, its Coding Index at 71.8, and its price at $2.00 per million input tokens and $6.00 per million output tokens.

What a "release" actually looks like for an open-weights model

A date question on a closed model is simple — the API flips on. For an open-weights model, "released" has layers, and each layer lands days apart:

• The raw weights appear as a Hugging Face and ModelScope repository, with a model card and a LICENSE file.

• Inference frameworks follow — vLLM and SGLang are expected to add support within days, which is the signal that a model is ready to serve rather than just download.

• Community quantizations (GGUF, AWQ) typically lag by one to two weeks, which is when Ollama, LM Studio, and llama.cpp users can actually pull and run it.

So if the question is "when can I download the weights," the answer is a repository appearing on the official Qwe​n org. If the question is "when can I run it with one command," add one to two weeks. The August 3 promise targeted the first of those for the week of August 10 — and that layer has now landed for one member of the generation: Qwen3.8-2.4T-A95B went up on Hugging Face that week. For the 27B specifically, it still has not landed on Hugging Face — but the ModelScope countdown now names a target: August 15, 2026 at 00:00 JST.

What is still unverified

Everything else about the 27B is speculative until the repository drops. Do not let confident coverage convert projections into specs — none of the following are confirmed at the time of writing:

• Architecture — dense or mixture-of-experts, and the active-parameter count if MoE.

• Context window, output limits, and whether it is text-only or multimodal. The ModelScope teaser, per PC Watch, describes native multimodal and multilingual support for images and video — vendor-stated on the teaser, still unverified until the weights drop.

• The license. Apache 2.0 is not a safe assumption; recent Qwe​n open releases shipped under the Tongyi Qianwen license, which carries a 100-million-monthly-active-users clause that triggers commercial discussion beyond that threshold.

• Any benchmark scores. Every figure circulating for the 3.8 generation is vendor-reported for the Max; the 27B has no public evals at all.

• Exact quantized sizes. The VRAM figures in circulation are projections carried over from Qwen3.6-27B, not measurements of the 3.8 model.

Why the license matters more than the date

The drop date determines when you can start testing. The license determines whether you can ship. A Qwe​n model under the Tongyi Qianwen license is free to use below 100 million monthly active users — past that, you are expected to discuss commercial terms. For a local model that teams will wire into products and agents, that threshold can arrive faster than it sounds. The first file to open in the repository is the LICENSE, before the model card and before the README.

How to catch the drop before the blog posts do

The official Qwe​n organization on Hugging Face is the ground truth — when a repository named Qwen3.8-27B appears there, or on ModelScope, the release has happened. The mechanism already proved itself this week: Qwen3.8-2.4T-A95B and its FP8 sibling appeared on the org exactly that way, so the 27B repo will be just as unmistakable. On ModelScope, a countdown page for Qwen3.8-27B is already live, pointing to midnight on August 15 — the closest thing to an official date yet. Three earlier signals are worth watching:

• A vLLM or SGLang pull request naming Qwen3.8-27B. Framework PRs are often the first public trace of a weight drop, because integration work starts the moment the repo is staged.

• The Qwe​n team's own channels (X, WeChat, the qwen.ai blog), which tend to announce the drop itself.

• Community quantizers — Unsloth and others usually post quant files and real VRAM measurements within days, which converts the projections above into measurable numbers.

One warning applies the whole time: before the official org publishes, any "Qwen3.8-27B" you can download is either a placeholder or a third-party upload. The name alone is not authenticity.

What you can use today

The 27B is the model to wait for, not the only option from this generation. Qwen3.8-Max is live right now — and self-hostable too, since its weights are downloadable — and on OrcaRouter it is billed at provider list price, $2.00 per million input tokens and $6.00 per million output tokens, with zero markup, so the price on the vendor's own pricing page is the price you pay. One key covers it, and when the 27B drops and earns trust on public benchmarks, the same key will route to it too. For a team that wants to build against the generation today while keeping the option of a self-hosted 27B later, that is the least-locked-in path there is.

The straightforward summary: the release-date answer is clearer than it was a week ago. The window Alibaba gave — the week of August 10, 2026 — came and went. It delivered the flagship's weights: Qwen3.8-2.4T-A95B and its official FP8 quantization are on Hugging Face now. It did not deliver the 27B's — but the delay now has a target: the ModelScope countdown runs to midnight on August 15, 2026 (00:00 JST, per PC Watch). Watch the official Qwe​n org on Hugging Face and the Qwe​n page on ModelScope, read the LICENSE the moment the 27B repo lands, and treat every spec and VRAM number circulating as a projection until the weights make them measurable.