Qwen3.8-Max Is 2.4T: What Open Weights Would Cost You (2026)

Alibaba previewed Qwen3.8-Max on July 19, 2026, a 2.4 trillion parameter multimodal model it describes as second only to Fable 5, and said open weights are coming with no date attached (MarkTechPost, 2026). Eight days later Moonshot actually shipped a 2.8T open-weight model, so the hardware question stopped being theoretical. Measured answer: expect 490GB to 800GB of weights at aggressive quantization, which means no Mac and no GPU, and a server that costs more than the laptop you are reading this on.

What is Qwen3.8-Max-Preview?

Qwen3.8-Max-Preview is Alibaba's new flagship, introduced at the World Artificial Intelligence Conference in Shanghai. The published specification is a 2.4 trillion parameter sparse mixture of experts, multimodal across text, images, video and documents, with a 1 million token context window. It is the first Qwen model above one trillion parameters to handle multimodal input.

Access today runs through Alibaba's Token Plan subscription, offered at 10% of standard pricing during the preview. That is a cloud API, not a download.

What Alibaba has not published

This section matters more than the specification, so it goes early. As of July 28, 2026, Alibaba has not released a model card, a benchmark table, the activated parameter count, a license, or technical documentation for Qwen3.8-Max-Preview.

That leaves the headline claim unverifiable. Second only to Fable 5 is Alibaba's positioning statement, not a third-party result, and there is no scored leaderboard entry to check it against. ModelFit does not publish numbers we cannot trace to a source containing them, so you will not find a Qwen3.8-Max row on our leaderboard until one exists to verify.

Open weights promised soon, with no date and no license, would also break Alibaba's pattern: the Max tier has historically stayed closed while the smaller Qwen models shipped openly. Treat it as an intention, not a commitment.

What would it take to run a 2.4T model?

Here the guesswork ends, because a comparable model shipped last week. Moonshot published Kimi K3, 2.8 trillion parameters, on July 27. We measured every published build from the HuggingFace API:

Scaling those ratios down to 2.4 trillion parameters gives a usable projection for Qwen3.8-Max, assuming a comparable quantization pipeline. Around 800GB at 2-bit, and around 490GB at the most aggressive 1-bit build. These are projections from measured ratios, not measurements of a model nobody outside Alibaba has.

Set that against what a single machine holds:

The Mac Studio line is the one to sit with. A 512GB Apple machine, the largest you can buy, would only fit a 2.4T model at the most degraded quantization level available, with almost nothing left for context. Kimi K3 missed that same ceiling by 55GB. Qwen3.8-Max would scrape past it and then run out of room for the KV cache.

For the actual buildable answer, including a priced parts list, see what it takes to run Kimi K3. The short version is roughly $7,300 of used dual-socket EPYC hardware with 1TB of DDR4, where memory is 68% of the bill.

Why the missing number is the important one

Total parameters set your memory bill. Activated parameters set your speed. Alibaba has published the first and withheld the second, which means nobody outside the company can estimate throughput.

The arithmetic is simple enough to show. Decode speed on a sparse mixture of experts is bandwidth-bound, and you read only the active parameters each token:

read per token = active params x bytes per parameter
tok/s ceiling  = memory bandwidth / read per token

Kimi K3 discloses 104B activated, so at its 2-bit build that is about 34.7GB read per token, and on a 341GB/s dual-socket box the ceiling lands near 9 tok/s est. Apply the same math to Qwen3.8-Max and you get nothing, because the first term is blank. A 2.4T model with 20B active and a 2.4T model with 150B active have the same memory bill and a sevenfold difference in speed.

So when the weights land, the activated parameter count is the first line to read. It decides whether the model is usable on a RAM box or merely loadable.

Should you wait for it?

No, in the sense that matters: do not defer a hardware purchase or a workflow decision on an undated promise from a company whose top tier has always stayed closed.

Two things could change that. If Alibaba ships weights with a permissive license, it becomes the largest openly licensed multimodal model available, which is a genuine event. And if the activated parameter count is low, a 2.4T model could be more practical to serve than its size suggests, the same way Qwen's smaller MoE models punch above their memory footprint.

Meanwhile the useful move is smaller models that already run. Qwen3.6-27B scores 77.2 on SWE-bench Verified on the independent vals.ai leaderboard (vals.ai, 2026) and fits a 32GB Mac, and a new 9B coding specialist now reports numbers close to models four times its size: Ornith 1.0 9B. For your specific machine, the ModelFit wizard or npx @wecko-ai/modelfit names a pick, and the RAM guide explains the sizing.

Why do 2T-plus open models keep appearing?

Because the competitive pressure runs between labs, not between labs and your laptop. Alibaba previewed Qwen3.8-Max days before Moonshot shipped Kimi K3's weights, and both are chasing the same frontier position. Neither release was designed for consumer hardware.

The part that reaches you is downstream: pruned variants, distillations, and the training recipes. One K3 experiment already cuts 80% of the experts to reach 350GB. If that kind of surgery preserves quality, the distance between a 2.4T release and something a workstation runs gets much shorter. That is the pipeline worth watching, and it lands in our open dataset when it produces something you can actually install.

FAQ

Can you run Qwen3.8-Max locally?

Not today. It is a cloud preview available through Alibaba's Token Plan subscription, with no weights published. Alibaba says open weights are coming but has given no date and no license. If they ship at 2.4 trillion parameters, expect roughly 490GB to 800GB of weights depending on quantization, which rules out every Mac and every consumer GPU.

How much RAM would Qwen3.8-Max need?

Projecting from measured Kimi K3 builds at 2.8 trillion parameters, a 2.4T model would land near 800GB at 2-bit and near 490GB at 1-bit. A 1TB server handles either. A 512GB Mac Studio would only fit the most degraded build, with little room left for context. No official figure exists because Alibaba has not published a model card.

Is Qwen3.8-Max open source?

Not yet, and possibly not ever in the full sense. Alibaba has promised open weights without a date, and has not published a license. Even if the weights ship, open weights and open source are different things: training data and methodology are rarely included.

How does Qwen3.8-Max compare to Kimi K3?

Kimi K3 shipped weights, Qwen3.8-Max has not. K3 is 2.8 trillion parameters with 104 billion activated and a published model card; Qwen3.8-Max is 2.4 trillion with the activated count undisclosed. On claims alone Alibaba positions its model as second only to Fable 5, but with no benchmark table published there is nothing to verify against K3's independently scored result.

What should I run instead right now?

Something that fits your machine today. Qwen3.6-27B is the strongest independently scored open coding model at 77.2 on SWE-bench Verified and runs on a 32GB Mac. On 16GB, Ornith 1.0 9B is a 5.6GB coding specialist reporting scores close to 35B models. Run npx @wecko-ai/modelfit for a pick matched to your hardware.

Sources

• MarkTechPost, Qwen3.8-Max preview report: https://www.marktechpost.com/2026/07/19/alibaba-previews-qwen3-8-max-a-2-4-trillion-parameter-multimodal-model-days-after-moonshots-kimi-k3-open-weight-launch/
• Kimi K3 model card, for the 2.8T comparison: https://huggingface.co/moonshotai/Kimi-K3
• HuggingFace model API, Kimi K3 file sizes measured July 28, 2026: https://huggingface.co/api/models/moonshotai/Kimi-K3
• vals.ai SWE-Bench Verified leaderboard: https://www.vals.ai/benchmarks/swebench