Qwen 3.8 vs Qwen 3.8-Max: What's the Difference and Which One Do You Need? (2026)

Alibaba shipped two models under one name. Sorting out which is which decides how you deploy.

If you searched "Qwen 3.8," you've probably noticed the coverage doesn't agree on what it is. Some articles describe a 2.4 trillion parameter frontier model. Others talk about running it on a single GPU. Both are right, because Qwen 3.8 is not one model.

Alibaba's August 3 launch covered two very different releases sharing a generation name: Qwen 3.8-Max, the API-only flagship, and Qwen3.8-27B, the small open-weight model promised within about a week of launch. Which one you should care about depends entirely on how you plan to use it. Here is the clean split.

TL;DR

• Qwen 3.8-Max: 2.4 trillion parameters (sparse MoE, ~95B active), 1M token context, API-only, $2 per million input tokens and $6 per million output. Launched August 3, 2026
• Qwen3.8-27B: the self-hostable release, out now under Apache 2.0. 28B dense with a vision encoder and 262k native context, and it fits on a single GPU
• The 27B shipped with a model card and benchmark numbers. The Max still has no official benchmark table, so its rankings remain vendor claims
• The 27B's weights are out: released August 13-14 on Hugging Face under Apache 2.0, with a surprise vision encoder. The Max-class weights have since landed as Qwen3.8-2.4T-A95B: text-only, custom license, 400GB+ even at 1-bit, so it changes little for most teams
• If you want capability through an API with zero infrastructure: 3.8-Max. If you want weights you control on hardware you choose: the 27B is downloadable today
• Running Qwen in production today: Qwen 3.8-Max and Qwen3.8-27B are live on Yotta AI Gateway

Why the Confusion Exists

Alibaba previewed Qwen 3.8-Max at WAIC in July, launched it officially on August 3, and announced the 27B alongside it. Most headlines just said "Qwen 3.8." The result is a name that means the flagship in one article and the small model in the next.

The naming pattern helps once you know it. In the Qwen lineup, "Max" has always meant the closed flagship tier: Qwen 3.7-Max was API-only, and 3.8-Max continues that. The open-weight line runs separately at smaller sizes, and Qwen3.8-27B is that line's next release, succeeding Qwen3.6-27B, one of the most-liked local models of the year.

So when someone asks "is Qwen 3.8 open source," the honest answer is: which one? The flagship API is closed, and its open counterpart turned out to be a stripped text-only checkpoint. The 27B was announced as open-weight from the start. As of August 13-14, the 27B's weights are live on Hugging Face; the flagship's open release exists too, but as a text-only 2.4T checkpoint most teams can't run.

Qwen 3.8-Max: The Flagship

Qwen 3.8-Max is the model behind the "second only to Claude Fable 5" headlines, and that framing is Alibaba's own internal evaluation, not a published benchmark. What is confirmed:

• 2.4 trillion total parameters, sparse Mixture-of-Experts, roughly 95B active per token (about 4%)
• 1M token context window, 128k max output
• Multimodal, with text plus visual inputs confirmed
• Standard API pricing: $2 per million input tokens, $6 per million output, $0.25 cached

That pricing is aggressive for a frontier model. For comparison, Moonshot's Kimi K3 lists at $3 in and $15 out. If the capability claims hold up, 3.8-Max undercuts its closest open rival on served price while being a fraction of the size.

What you cannot do with 3.8-Max is hold it. The Max API model itself still has no weights, no fine-tuning, and no self-hosting. The closest thing is Qwen3.8-2.4T-A95B, an open text-only checkpoint: no vision, no native 1M context, a custom license, and a 400GB+ footprint even at 1-bit, which keeps it out of reach for all but cluster-scale teams. You consume it as a service and your costs scale with tokens, not hardware. Our Qwen 3.8 API access guide covers the access paths and pricing in detail.

Qwen3.8-27B: The One You'll Actually Run

Nobody self-hosts a 2.4T model. A 27B is a different story, and that is why the 27B may end up the more consequential half of this launch for infrastructure teams.

The release delivered more than promised: 28B parameters, dense with hybrid attention, a vision encoder (the release is multimodal), 262k native context extensible to 1M, and an Apache 2.0 license, all live at the official Qwen/Qwen3.8-27B repo, with quantized community builds already listed for Ollama and llama.cpp.

The confirmed math: roughly 56GB of VRAM at BF16, around 28GB at FP8, and 14 to 17GB at 4-bit, plus KV cache on top. That means everything from a rented RTX 4090 up to an H100 depending on precision and workload. The full memory breakdown, GPU options, and serving plan are in our Qwen 3.8 27B hardware guide.

Qwen 3.8-Max vs Qwen3.8-27B

Which One Should You Use?

The decision is less about which model is better and more about which failure mode you can live with.

Choose 3.8-Max when the hardest queries justify frontier capability and you don't want to touch infrastructure. You get the flagship's full capability, 1M context, and multimodality for a metered price. The tradeoff: no control, no fine-tuning, and economics that scale linearly with usage forever.

Choose the 27B when volume, cost control, or data locality matter. Fixed hardware cost, weights you hold, fine-tuning rights. The tradeoff: a 27B will not match a 2.4T flagship on hard reasoning, and you own the serving stack. Our vLLM vs SGLang comparison covers the engine choice when you get there.

Most production teams end up splitting: the big model through an API for the hardest 10 percent of traffic, a small self-hosted model for the volume work. The Qwen 3.8 generation is clearly built to serve exactly that split from one family.

One honest caveat before committing to either: with no published benchmarks, nobody knows yet how much capability the API premium buys over the 27B. That difference is the number to watch when scores land. We track what's verified in our Qwen 3.8 benchmarks post.

And if you're cross-shopping outside the Qwen family, the most common matchup is against the open-weight incumbent: our Qwen 3.8 vs GLM 5.2 comparison covers that head-to-head.

Where Qwen 3.7-Max Fits

One more model gets mixed into "qwen 3.8" searches: Qwen 3.7-Max, the previous flagship. It is still the Qwen you can put into production through a router today, with published, third-party-comparable benchmarks, which 3.8-Max still lacks.

Qwen 3.8-Max and Qwen3.8-27B are both live on Yotta AI Gateway with OpenAI-compatible endpoints, one API key alongside DeepSeek, GLM, Kimi, and the rest of the catalog, with Claude being added. Switching between them is a config change. That is the practical hedge while the 3.8 generation's claims get verified: build on 3.7-Max now, evaluate 3.8 as the evidence arrives. The deployment logic is in how to run Qwen 3.7 in production.

Frequently Asked Questions

Is Qwen 3.8 the same as Qwen 3.8-Max?

No. Qwen 3.8-Max is the 2.4T API-only flagship. "Qwen 3.8" also covers Qwen3.8-27B, the open-weight small model announced at the same launch. Coverage uses the names loosely.

Is Qwen 3.8 open source?

Qwen 3.8-Max as a product remains API-only. A text-only 2.4T checkpoint, Qwen3.8-2.4T-A95B, is on Hugging Face under a custom license, but it drops the vision and native 1M context that define the Max.

What does Qwen 3.8 cost?

Qwen 3.8-Max lists at $2 per million input tokens, $6 per million output, $0.25 cached, through Alibaba's standard API. The 27B is free to download under Apache 2.0; you pay for the hardware you run it on.

Can I run Qwen 3.8 locally?

Yes, the 27B, as of its August 13-14 release. Quantized builds run on roughly 17GB, which puts a 24GB consumer GPU in play; our local setup guide covers the fastest path. The 2.4T checkpoint is not a local option; it needs cluster-scale hardware.

Is there a Qwen 3.8 27B?

Yes, and it's out: released August 13-14 under Apache 2.0, with a vision encoder and 262k context. Our hardware guide covers the confirmed specs and memory math.

Is Qwen 3.8-Max better than Kimi K3?

Unknown. Neither ranking can be verified: 3.8-Max has no published benchmark table, and head-to-head comparisons don't exist yet. Our Qwen 3.8 vs Kimi K3 breakdown covers what evidence there is.

Which Qwen model can I use in production right now?

Qwen 3.8-Max and Qwen3.8-27B are live on Yotta AI Gateway, alongside 3.7-Max with its verified benchmarks and published pricing.

Bottom Line

"Qwen 3.8" is a family, not a model. The Max is a 2.4T flagship you rent by the token. The 27B is the one you download and run on a single GPU, available now. Independent benchmarks are still pending for both.

The practical move: if you need frontier capability through an API today, 3.8-Max is live and priced aggressively. If you self-host, the 27B is live and the local setup guide gets you running. Both are one API key away on Yotta AI Gateway.