Qwen3.8-Max open weights (2.4T-A95B): what it takes to run, and the 27B you can actually run

Alibaba did what it promised. The Max-class Qwen3.8 model now has public weights, under the name Qwen3.8-2.4T-A95B. The model card says it plainly: "For the first time, Qwen3.8 brings a Qwen-Max-class model to open release."

I read the model card, the config file and the file lists on Hugging Face on 4 October 2026. Then I ran the numbers through our VRAM engine. I have not loaded this model on anything. What follows is arithmetic on published files, which is the honest way to answer "can I run it?"

Are the Qwen3.8-Max weights actually open?

Yes. The weights live at Qwen/Qwen3.8-2.4T-A95B on Hugging Face, created on 8 August 2026, with an official FP8 version next to it. The open model is text-only and always thinks before it answers. The hosted Qwen3.8-Max adds image input, a non-thinking mode and a 1M-token default context.

That split matters if you compare the two. The model card describes Qwen3.8-Max as "the official version based on Qwen3.8-2.4T-A95B with more features." So the API you pay for and the files you download are related, but not identical.

The license is custom, not Apache-2.0 like the smaller Qwen3.8 27B. It grants broad rights to use, modify, host and fine-tune. Two conditions bite only at scale:

• A product with more than 100 million monthly users or more than $20 million a month in revenue must display the model name in its interface.
• A model-as-a-service or AI work assistant business whose group revenue passes $50 million in any 12 months needs a separate license from Qwen.

For a homelab, a startup or a research group, neither applies. Read the LICENSE file yourself before you build a business on it. This is a summary, not legal advice.

How much memory does Qwen3.8-2.4T-A95B need?

About 1.3 TB just for the weights at a 4-bit-class quant, and 2.5 TB at the official FP8. Even the smallest community quant, at about 1.3 bits per weight, is 397 GB. Add the context cache and runtime on top. Every figure here is far beyond any consumer GPU, or several of them.

These are the real file totals, summed from each repo's file list on 4 October 2026.

The bits-per-weight column is derived: file size times 8, divided by the 2,446,182,725,504 parameters Hugging Face reports. Our engine's Q8_0 estimate for this model is 2,421 GiB, within a gigabyte of the real Q8_0 file. That gives me some confidence in the rest of the math.

Why doesn't "95B active" make it smaller?

Because active parameters set the compute per token, not the memory. The router picks 10 of 512 experts for every token, plus one shared expert. Any expert can be next, so all of them must sit in fast memory. You pay memory for 2.4 trillion parameters and compute for about 95 billion.

That is the same trap every giant mixture-of-experts model sets. The upside is real: if you can hold it, it generates tokens far faster than a 2.4T dense model would. The n-cpu-moe offload trick helps when a model is a little too big for your GPU. It does not help when the model is a hundred times too big for your whole PC.

One thing is cheaper than you'd expect: the context cache. The config shows 92 layers, but only 23 use full attention and keep a cache. The rest use Gated DeltaNet linear attention with a small fixed state. With 4 KV heads of 256 dimensions, our engine puts the fp16 cache at about 92 KiB per token, or about 23 GiB at the full 262,144-token native window. Next to 1.3 TB of weights, that is a rounding error.

Can I run Qwen3.8 on a single GPU or a Tesla P40?

Not the 2.4T model. Its smallest quant is 397 GB. But Qwen3.8 27B, a separate dense model published on Hugging Face on 5 August 2026 under Apache-2.0, fits one 24 GB card. Our engine puts it at about 18.5 GB at Q4_K_M with 32K tokens of context. A 24 GB Tesla P40 has room for that.

The 27B has 27,781,427,952 parameters and the same hybrid layout: only 16 of its 64 layers keep a cache. That keeps long context cheap. Here is what our engine says, in GiB, with an fp16 cache and our standard 0.8 GiB runtime allowance.

So on a 24 GB card like the P40 or an RTX 3090, Q4_K_M with 32K context fits with room to spare, and Q5_K_M fits too. The 128K budget does not. These are memory numbers only. The P40 is an old Pascal card, so check the software side in our Tesla P40 guide before you buy one for this. For a card-by-card view, see best GPU for Qwen3.8 27B and its cost to run page.

If you have only 12 or 16 GB, a smaller MoE model is the better fit. Our list of the best MoE models for 12 to 16 GB GPUs covers the ones that run there.

How does Qwen3.8-Max score against GPT-5.6 and Claude?

Close to the frontier, not ahead across the board. On Alibaba's own model-card table it leads on PaperBench (93.0) and IFBench (82.8). It trails Claude Fable 5 on SWE-bench Pro (67.7 vs 80.0) and GPT-5.6 Sol on Terminal Bench 2.1 (86.6 vs 88.8). These are vendor numbers, not independent tests.

Read the footnotes before you quote these. Alibaba ran several tests in the Claude Code harness and took other labs' best published scores from different harnesses. It also notes that Fable 5 results "may involve fallbacks." Fair enough as a vendor table. Not the same as a head-to-head lab test.

Should you rent it or use the API?

Use the API unless you run a GPU cluster. Qwen Cloud lists Qwen3.8-Max at $2 per million input tokens and $6 per million output tokens (read 4 Oct 2026). Self-hosting only pays off for teams that already own racks of datacenter GPUs and need the weights in-house.

That is who this release serves: cloud providers, research labs and big companies that want a Max-class model without per-token fees. It is the same pattern as Kimi K3's open-weight launch: open in license, out of reach in hardware. The price pressure is real too, and it lands next to moves like OpenAI's 80% GPT-5.6 price cut.

The useful news for home builders is not the 2.4T model. It is that the same generation ships a 27B that fits one 24 GB card with long context to spare. Size your own card in the VRAM calculator, and leave the 2.4T one to the people with racks.