Qwen3.8-27B open weights under Apache 2.0, and the 17GB footnote
Alibaba released Qwen3.8-27B under Apache 2.0: 27 billion parameters, a 262K token context, and roughly 17GB to run it locally. That last number is the answer to the frontier-on-a-phone headline.
Alibaba released the weights for Qwen3.8-27B on August 14 under an Apache 2.0 licence: a dense 27 billion parameter model with a native 262K token context that stretches to 1M. The Qwen team says running it locally takes about 17GB. That single number is what the frontier-on-a-phone framing has to survive, and it mostly does not.
what actually shipped
Alibaba Cloud put out the files developers and hardware makers need to host the new multimodal series on private infrastructure. The 27B dense model is tuned for real coding and office workflows, and it arrived beside open weights for the much larger Qwen3.8 2.4T A95B, which we covered when the 2.4 trillion parameter Max weights went out.
MediaTek claims Day 0 support for the 27B model on its flagship Dimensity mobile SoC and on the Auto Cockpit C-X1, aimed at agents running inside phones and cars. The remaining partners cover the three other places weights get executed: LM Studio for laptops, Nvidia RTX Spark for local development, and Cerebras for dedicated servers, with a Shared Tier slot promised soon.
Friday's version of this story ran a different headline, that Qwen3.8-27B beats Claude Opus 4.6 Max on agent benchmarks. Nothing in the material behind it carries a number: no table, no benchmark name, no score. Reading a model benchmark table properly is written for exactly this gap, where the claim exists and the column does not.
what it changes for you
Open weights are the only form of a serious model that reaches a user without an account, a card or a payment rail. Once the file sits on your own disk, inference never touches an API, so there is no layer left for anyone to close. The catch is one step earlier: offline inference still needs one online download, and the hosts are Hugging Face and Alibaba Cloud, both of which have gates of their own. Being offline in step two does not undo being online in step one.
The 17GB figure is worth doing the arithmetic on. Twenty-seven billion parameters at 16-bit precision would be roughly 54GB of raw weights, so 17GB means compression. The release does not name the quantisation level, does not publish tokens per second, and does not say at what context length the measurement was taken, while the memory a 262K token window eats sits on top of that number rather than inside it. A laptop with enough RAM can hold this. The phone already in someone's pocket cannot, and every open model we have tracked in our AI release coverage has had the same distance between the demo rig and the shipping device.
the case against
A model that fits in 17GB is not a frontier model in the strict sense, and the release itself is the proof. Alibaba opened the 2.4T A95B weights on the same day. If the small one did the same job, the large one would have no reason to exist.
Day 0 support is a chip vendor's statement, not a report of the model running on a handset anyone can buy. The part of the story the headline rests on is the one part published without a measurable figure, which is the shape the Grok 4.6 launch took when the numbers arrived.
Apache 2.0 is the strongest thing in this announcement, because unlike the bespoke licences some labs still file under the words open weight, it carries no user ceiling and no company-size condition. Even so, the licence text living on the repository is not reproduced here, and the LICENSE file is the only version that binds a commercial deployment. One number would settle the rest of it: tokens per second on a real Dimensity, measured by someone Alibaba did not build a stage demo for. 🔍
Frequently asked questions
Does Qwen3.8-27B actually run on a smartphone?
Nobody has shown it running on a shipping handset yet. What was announced is MediaTek Day 0 support on the flagship Dimensity SoC and the Auto Cockpit C-X1, which is chip readiness rather than a measured phone run. The launch material names no device, no handset RAM figure and no tokens per second. The only official number is the roughly 17GB local footprint.
How much memory does Qwen3.8-27B need locally?
About 17GB, according to the Qwen team. That figure describes compressed weights, since 27 billion parameters at 16-bit precision would come to roughly 54GB. The release does not say which quantisation level produces the 17GB, or at what context length it was measured, and the memory a 262K token window consumes sits on top of that number rather than inside it.
Is commercial use of Qwen3.8-27B allowed?
The stated licence is Apache 2.0, which permits commercial use, modification and redistribution with no user cap or company-size condition attached. The full licence file sitting on the repository is not reproduced in this launch material, so reading the LICENSE file in the repo itself remains the only reliable check before any commercial deployment.
