FRI, SEPTEMBER 04, 2026
Independent · In‑Depth · Practitioner‑Tested
✎ Large Language Models

GPU Memory Is the Expensive Constraint. Alibaba Is Testing a Way Around It

Qwen3.8-Flash-Next is an open-weight release previewing the architecture Alibaba plans for Qwen4, reported at 125 billion total parameters with roughly 6 billion active per token — plus a separate 51 billion parameter component designed to sit in ordinary system memory rather than GPU memory. If that holds up under real latency testing, the hardware requirement for local models changes shape.

By AIToolsRecap September 4, 2026 6 min read 38 views
Home Articles Large Language Models Qwen Alibaba Put Part of a Model in System RAM — and...
WHAT WAS RELEASED

● Qwen3.8-Flash-Next, open weights, published to preview the architecture planned for Qwen4.

● Reported size: 125 billion total parameters, roughly 6 billion active per token.

● The unusual part: a separate 51 billion parameter component designed to run in system RAM rather than GPU memory.

● Caveat: figures are as reported and not independently verified.

Why the memory split matters

The constraint on running a large model locally is almost never raw compute. It is GPU memory — VRAM is expensive, fixed at purchase, and the thing that decides whether a model fits on your hardware at all.

System RAM is cheap by comparison. A workstation with 128GB of ordinary memory costs a fraction of a card with equivalent VRAM.

THE ARCHITECTURAL BET

If a meaningful chunk of a model can live in system memory without destroying latency, the hardware requirement changes shape entirely. You stop buying VRAM and start buying RAM, and the two are not remotely comparable in price.

Whether it works in practice is the open question, and it is one the community will answer within days now the weights are public.

The sparsity is aggressive

125 billion total parameters with roughly 6 billion active per token is a ratio of about twenty to one. Mixture-of-experts designs have been moving this direction for two years, but that is sparse even by current standards.

For comparison, GLM-5.3 runs roughly 743 billion total with about 40 billion active — a ratio near nineteen to one but at six times the scale. Kimi K3 and DeepSeek sit in similar territory on ratio.

The practical consequence of high sparsity is that inference cost tracks active parameters while capability tracks total. That is the whole reason the architecture is popular, and pushing the ratio further is the obvious next move.

What a preview release is for

Alibaba published this specifically to preview Qwen4 architecture rather than as a flagship. That is a deliberate strategy and worth understanding.

What it gets Alibaba Detail
Free testing at scale Thousands of people will stress the memory design in ways an internal team cannot
Tooling readiness Inference engines can add support before the flagship lands
Mindshare ahead of Qwen4 Developers who tune workflows to this architecture are ready buyers later
Low downside A preview that underperforms costs reputation on a preview, not on a flagship

Check the licence before you build on it

Open weights and permissive licensing are not the same thing, and the gap has widened this year. Alibaba shipped Qwen 3.8-Max under a custom licence rather than Apache. Kimi K3 ships under a modified MIT. Z.ai held GLM-5.3 weights for two weeks and has still not stated the flagship licence.

Downloadable tells you nothing about what you are permitted to do with the output, whether commercial use is unrestricted, or whether the terms can change. Read the actual file rather than the announcement.

What to do with it

If you are... The read
Running models locally This is the release to watch. Wait a week for community benchmarks on real hardware
Choosing a production model Not this one. Preview releases are for learning, not for depending on
Budgeting hardware Do not buy on this yet. Wait until someone measures latency with the RAM component in use
Tracking Chinese labs Publishing architecture ahead of a flagship is a confidence signal worth noting

Sources

FAQ

What is Qwen3.8-Flash-Next?

An open-weight model from Alibaba's Qwen team, released to preview the architecture planned for the next-generation Qwen4 models rather than as a flagship in its own right.

How big is it?

Reported at 125 billion total parameters with roughly 6 billion active per token, plus a separate 51 billion parameter component intended to run in system memory. These figures are as reported and not independently verified.

Why put part of a model in system RAM?

GPU memory is the expensive constraint on running large models locally. System RAM costs a fraction as much, so shifting part of the model there changes what hardware is required — if latency holds up.

Can I use it commercially?

Check the licence file rather than assuming. Alibaba shipped Qwen 3.8-Max under a custom licence rather than Apache, and open weights do not imply permissive terms.

Should I use it in production?

No. It is explicitly a preview of architecture. Use it to learn what Qwen4 will look like, not to build on.

When is Qwen4 expected?

Not announced. This release previews the architecture, which usually means a flagship follows within months rather than weeks.

Tags
AlibabaQwenOpen SourceModel ReleaseHugging FaceLocal AILicensing2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →