Is Qwen3.8-Max open source?

Not at launch. Alibaba announced it on August 3, 2026 and said the open weights would follow the next week — its first open-source model at this scale. Until then it's API-only, and the license is still unpublished.

How big is Qwen3.8-Max?

2.4 trillion parameters total, with 95 billion active per token (a mixture-of-experts design), and a 1-million-token context window.

qwen3.8-max: specs, pricing, and how to access it

Q: How much does Qwen3.8-Max cost?

$2 per million input tokens, $6 per million output, and $0.25 per million cached input, via the QwenCloud API.

Qwen3.8-Max is Alibaba’s biggest model yet: 2.4 trillion parameters, 95 billion active at once, a 1-million-token context window, multimodal. You can rent it today through QwenCloud — $2 per million input tokens, $6 per million output. You can’t download it yet; the open weights land next week, if the license cooperates.

how much does qwen3.8-max cost?

$2 per million input tokens, $6 per million output, $0.25 per million cached — QwenCloud API only. Cheaper on output than the US frontier, dearer than the budget Chinese models. Mid-tier money for something Alibaba is selling as top-shelf. There’s no self-host option yet, so that API bill is the whole price.

can you download qwen3.8-max? (open weights)

Not yet. It was announced August 3; the weights are promised “next week” — the first time Alibaba open-sources anything this big. Until they show up, “open-source” is a headline, not a file you can run. Building a plan around local inference or fine-tuning? That starts next week at the earliest, and only if the license lets you — which Alibaba hasn’t published either. Two unknowns stacked on a launch date.

how good is qwen3.8-max, really?

Alibaba says it beats Anthropic’s Fable 5. On benchmarks Alibaba ran. Grade your own homework and you’ll ace it too — self-reported numbers are marketing until someone independent reruns them. The one score that isn’t Alibaba’s: the crowdsourced Arena.AI leaderboard, where it’s the top Chinese model for text and second in the world for vision, behind a Fable 5 variant. That’s real, and it’s genuinely strong. It’s also the only number here you should trust yet.

what is qwen3.8-max built for?

Mixture-of-experts on the Qwen3.5 architecture, pointed at coding, research, and professional work. The multimodal side is aimed at documents and screens — reading a PDF or a UI, not just producing text. The 1-million-token context is the real selling point there: hold a whole codebase or a document stack in one shot. Kimi K3 already does the same million tokens, so this is the price of entry at the top, not an edge.

how do you access qwen3.8-max right now?

QwenCloud’s API, and nothing else. Sign up, grab a key, pay per token. No download, no self-host, and the app front-end is the same API underneath. Want it on your own hardware? Set a reminder for the week of August 10 and hope the weights actually ship.

So here’s the real status: a huge, probably excellent model you can rent today and own never, whose report card was written by the company that built it. The Arena ranking says the talent is there. Next week says whether “open” was real or just a nicer word for “coming soon.” Right now it’s the biggest model you can’t download.