🧠 Qwen3.8-Max: What's Confirmed and What Isn't Yet (2026)

Alibaba shipped Qwen3.8-Max on August 3, and three days later I still can't put it on my rig. There's no download. What launched is hosted-API-only, and the open weights Alibaba promised for both Qwen3.8-Max and its smaller Qwen3.8-27B sibling aren't due until "next week" from the announcement, which puts a real release around August 10. Anyone posting local tokens/sec numbers on this model right now is either running the API through a wrapper and mislabeling it, or making it up.

That's the honest starting point, and it sets the format for the rest of this page: what Alibaba actually confirmed, what's vendor-claimed and untested, and what doesn't exist on a hard drive yet.

Confirmed and usable

Right now "usable" means the hosted API, full stop: OpenAI-compatible and DashScope-compatible endpoints through Alibaba Cloud Model Studio, plus access through the QwenWork workplace-agent platform. No local checkpoint exists.

The confirmed architecture: Qwen3.8-Max is a 2.4 trillion parameter Mixture-of-Experts model, with roughly 95 billion parameters activated per token, running on a hybrid attention mechanism. Context tops out near 1M tokens — precisely, max input around 991K tokens (983K with extended thinking switched on), max output 131K tokens, and a reasoning budget that alone can run to 262K tokens. Input takes text, image and video; output is text only.

The API tier is a real product, not a demo: 2M tokens per minute and 15K requests per minute of rate limit, function calling, structured outputs, batch processing, prefix completion, fine-tuning support, and five built-in tools including a code interpreter and web search/extraction.

Calling it "open" this week is getting ahead of the facts. What's actually shipped is a very capable, very closed API with an open-source promise attached.

Confirmed but vendor-run

Alibaba's own benchmark numbers: Terminal-Bench 2.1 at 86.6, GPQA Diamond at 92.6, PaperBench at 93.0, OSWorld-Verified at 86.1.

Real numbers, real announcement. I'm not waving them off. But they're Alibaba grading its own homework on Alibaba's chosen test set, with no independently reproduced run attached yet. Read the chart as a claim, not a verdict, until someone outside Alibaba runs the same evals and publishes the diff.

Third parties with API access are already circling the coding side, though. Testers are putting Qwen3.8-Max up against Kimi K3 on frontend, backend and agentic work within days of the launch, which at least says the API held up under real traffic.

None of that replaces a controlled benchmark run with a published methodology, but it's a faster signal than waiting on a formal eval, and right now it's the best read available before open weights make independent testing possible at all. I'd want three of those videos to agree before I trusted any single one.

Announced, not shipped

Open weights for Qwen3.8-Max. The first time Alibaba has open-sourced a Qwen-Max-class flagship, which is the actually interesting part of this launch if it holds. Weights promised "next week" from the announcement, so watch for something around August 10.

Open weights for Qwen3.8-27B. The smaller, dense sibling that's the one worth caring about on a single-GPU rig, following the naming pattern of last generation's Qwen3.6-27B. No exact date beyond "coming."

Pricing for anything besides the hosted API. Self-hosted licensing terms and fine-tuning cost on your own infrastructure aren't public yet. Any "total cost to self-host Qwen3.8-Max" post you see before the weights land is a guess dressed up as math, mine included if I'm ever tempted to run one early.

What a 2.4T MoE actually means for your GPU

This is where "billions of parameters" stops being a headline and starts being a memory problem. Qwen3.8-Max activates about 95B parameters per token, which sounds almost within reach until you remember what MoE inference actually requires: every expert has to sit in memory, because the router can send any token to any of them. It's less like booking one hotel room and more like renting every room in the building because you don't know which one the guest wants tonight. Active-parameter count doesn't shrink the hardware bill.

That's an enterprise-cluster number, not a GPU-tier one. For nearly everyone reading this, Qwen3.8-Max is a cloud-only proposition. My own 4080 isn't in the conversation, and neither is yours.

If you need Max-class output today, the API is priced by the token instead of by the card you don't have to buy: $2 in, $6 out, per million tokens. For most side projects that's cheaper than the power bill a 2.4T model would rack up running locally, even in the hypothetical world where you'd found the memory to load it. Self-hosting it yourself isn't a rig upgrade anyone reading a GPU guide is making — that's a company-sized infrastructure decision.

The 27B is the one to actually watch. I haven't run it, nobody outside Alibaba has, but the name and the size class echo Qwen3.6-27B closely enough to make a hedged comparison worthwhile: that model landed around 17GB of VRAM at Q4_K_M, fit a single RTX 4090 with room to spare, and pushed 35-55 tokens/sec on that card. None of those figures are confirmed for 3.8-27B. Dense parameter counts, attention design and context handling all shift release to release, and any of them can move the VRAM math in a direction the name alone won't tell you.

If it lands anywhere near that footprint, though, it's the same 24GB tier that's carried most of our local-LLM coverage all year, and I'll have real numbers within days of the weights dropping. Not vendor numbers. Mine, off my own hardware, using the same VRAM-first math that decides every GPU buy on this site.

If you're deciding whether to wait or run something today, our Qwen3.6-27B setup guide is still the current answer, and the quantization math that got that model onto a 24GB card will apply just as directly once 3.8-27B lands. Starting from zero on local inference? Run Llama locally first — the pull-and-go workflow barely changes between model families.

By the numbers

Item	Status	Figure
Announced	Confirmed	August 3, 2026
Access	Confirmed	Hosted API only (OpenAI- and DashScope-compatible)
Architecture	Confirmed	2.4T MoE, ~95B active params, hybrid attention
Max input context	Confirmed	~991K tokens (~983K with extended thinking)
Max output	Confirmed	131K tokens
Max reasoning budget	Confirmed	262K tokens
Rate limits	Confirmed	2M tokens/min, 15K requests/min
Input pricing	Confirmed	$2.00 / 1M tokens
Output pricing	Confirmed	$6.00 / 1M tokens
Cached input (implicit)	Confirmed	$0.25 / 1M tokens
Terminal-Bench 2.1	Vendor-run	86.6
GPQA Diamond	Vendor-run	92.6
Qwen3.8-Max open weights	Announced	"Next week," ~August 10
Qwen3.8-27B open weights	Announced	Coming, no date
VRAM for the 27B	Not yet confirmed	Comparison only: Qwen3.6-27B ran ~17GB at Q4_K_M

Every confirmed figure above traces to Alibaba's own Qwen3.8-Max announcement and MarkTechPost's launch-day coverage, both read this week.

I have zero tokens/sec, zero VRAM readings and zero quant sizes for anything in the Qwen3.8 family. That's not a gap in my testing, it's the state of the model: nobody outside Alibaba has weights to test yet. The day they land, I'm pulling the 27B onto the rig before I finish my coffee, and this page's numbers section stops borrowing from last generation.