Qwen 3.8 Max Explained: Alibaba's 2.4T Flagship Goes Open Weight

Update, August 14, 2026: the weights are out, and the licence is not what the precedent suggested. The Max checkpoint went up on Hugging Face on August 12 as Qwen/Qwen3.8-2.4T-A95B under Alibaba's own qwen3.8-max terms, which is not Apache 2.0, so read them before you build on it. The consumer sized Qwen 3.8 27B followed on August 13 and is Apache 2.0. If you came here to find out what you can actually run, that is the one: see How to Run Qwen 3.8 27B Locally for measured file sizes and the setup traps.

On August 3, 2026, Alibaba made Qwen 3.8 Max generally available, two weeks after teasing it at the World AI Conference in Shanghai. It is the largest model the Qwen team has ever shipped: roughly 2.4 trillion parameters, a 1 million token context window, image and video input, and launch benchmarks that put it at the top of the open ecosystem. The kicker for people like us: Alibaba says an open weight checkpoint follows within the week, and a much smaller Qwen 3.8 27B goes open weight at the same time.

Here is everything actually known so far, what is claim versus fact, and why the 27B sibling might matter more to you than the headline model.

What Qwen 3.8 Max Is

Qwen 3.8 Max is a sparse Mixture of Experts (MoE) model. The 2.4 trillion number is the total parameter count; launch coverage puts the active set around 95 billion per token, which is what keeps inference costs sane. It takes text, images, and video as input and answers in text, and it is positioned squarely at coding and agent workloads: function calling, structured outputs, and a set of built in tools (code interpreter, web search, web extractor, and image search) ship with the API.

The rollout started as a preview on July 19 at WAIC, ran through Alibaba's subscription platforms at a discount, and reached general availability today through OpenAI compatible and DashScope endpoints. The API also speaks the Anthropic request format, a first for the Qwen line.

The Spec Sheet

The Benchmark Claims (Read the Fine Print)

The launch numbers are strong: 92.6 on GPQA Diamond, 86.6 on Terminal-Bench 2.1, 86.1 on OSWorld-Verified, and huge generational jumps on the agentic coding suites (FrontierSWE 73.5, up from 40.7 for its predecessor; DeepSWE 1.1 at 56.6, up from 21.6). On the crowdsourced Arena.AI leaderboard it immediately became the highest ranking Chinese model for text.

Two honest caveats. First, these are launch day, vendor picked numbers, and launch week scores have a history of shrinking in the wash. Second, even Alibaba's own framing is "second only to Claude Fable 5", so this is a claim about leading the open pack, not the frontier. That claim, at least, the Arena.AI ranking already supports.

The direction matters more than the decimals. The gap between open ecosystem models and the closed frontier keeps narrowing, and it is the Chinese labs (Qwen, Moonshot, DeepSeek, Zhipu) doing most of the narrowing. Qwen 3.8 Max is the strongest entry in that series yet.

Pricing: Aggressive for a Flagship

The API costs $2 per million input tokens and $6 per million output, with cache reads from $0.17. That significantly undercuts Western flagships and even beats Kimi K3's $3/$15. Rate limits at launch are generous too: 2 million tokens per minute and 15,000 requests per minute. Fine tuning, batch jobs, and prefix completion are supported from day one.

For working code and every access route that exists today, see the companion piece: How to Run Qwen 3.8 Max.

The Open Weights Were the Real Event

Both checkpoints are public now, and the two halves came on different terms. The Max weights landed August 12 as Qwen/Qwen3.8-2.4T-A95B under a custom qwen3.8-max licence, so the Apache assumption in the original version of this post was wrong. The Qwen 3.8 27B landed August 13 under plain Apache 2.0, which does permit commercial hosting and fine tuning.

Why this matters more than launch day:

• Independent hosting. Once the Max weights are out, neutral inference providers can serve it, prices fall further, and your prompts no longer have to route through Alibaba's cloud.
• Real benchmarks. The community can verify the launch claims instead of taking screenshots on faith.
• The 27B is the local story. A current generation Qwen at 27B should run on a 16 GB GPU at 4 bit, exactly like today's Qwen 3.6 27B does. And weights mean derivatives: quantizations, distills, and eventually the abliterated and heretic builds that the uncensored ecosystem is built on. No weights, no freedom.

Until the drop, treat Qwen 3.8 Max as a very impressive hosted product with an open promise attached.

Can You Run It Locally?

The Max: no, and it is worth being straight about that. A 2.4T parameter model at 4 bit quantization is around 1.2 terabytes of weights before you allocate a single byte of KV cache for that 1M context. That is multi node datacenter hardware, not a gaming PC.

We did the full math, including what the 27B changes and what belongs on your GPU today, in a companion piece: Can You Run Qwen 3.8 Locally? The one sentence preview: the Max stays in the datacenter, the 27B lands on your desk, and until it does, Qwen 3.6 remains the best Qwen you can actually run.

LU Labs published the measured version of that math, so if you want the file size of every quant and the context cost per token, read how to run Qwen 3.8 27B locally: VRAM table and setup.

Wondering whether to wait for 3.8 or run 3.6 now? We lined them up: Qwen 3.8 vs Qwen 3.6.

What This Means for Local AI

Qwen is the family that open model people actually run. Its 3.6 generation powers some of the best uncensored local models of 2026, and its abliterated builds dominate the download charts. A new generation going open weight, with a consumer sized 27B in the same drop, is the best possible version of this news: frontier adjacent quality trickling straight down to hardware you own.

If you are new to local models, our beginner guide to running AI locally and the uncensored setup guide get you from zero to a working chat in minutes with Locally Uncensored.

FAQ

What is Qwen 3.8 Max?

Alibaba's flagship model, generally available since August 3, 2026: a Mixture of Experts model with roughly 2.4 trillion total parameters, a 1 million token context window, and text, image, and video input. It is the most capable model in the Qwen family to date.

Is Qwen 3.8 open source?

Partly. The weights are public, but only one of the two is under an open licence. Qwen 3.8 27B is Apache 2.0. The 2.4T Max checkpoint, published as Qwen/Qwen3.8-2.4T-A95B, carries Alibaba's own qwen3.8-max terms, so treat it as open weight rather than open source.

How much does the Qwen 3.8 Max API cost?

$2 per million input tokens and $6 per million output tokens, with cache reads from $0.17. Served through OpenAI compatible and DashScope endpoints.

Is Qwen 3.8 Max better than Claude or GPT?

It is the top ranked Chinese model on Arena.AI and trails Claude Fable 5. Alibaba's own claim is "second only to Fable 5". Independent verification of the launch scores is still thin.

Can I run Qwen 3.8 on my own computer?

Not the Max (~1.2 TB at 4 bit). The Qwen 3.8 27B going open weight alongside it is the one for your GPU. See our full hardware breakdown.

Getting Started with Local AI Today

Max sized models need a datacenter, but everything up to about 70B runs beautifully at home, uncensored, private, and free:

git clone https://github.com/PurpleDoubleD/locally-uncensored.git
cd locally-uncensored
# Windows: setup.bat | Linux: ./setup.sh

Or grab the installer from the releases page, open the Model Manager, and one click a model. When the 3.8 weights drop, we will cover the derivatives worth caring about.

Locally Uncensored is AGPL-3.0 licensed and free to use. Built by PurpleDoubleD.