Can You Run Qwen 3.8 Locally? The 2.4T Reality — and the 27B You Actually Can

Last updated: August 2026

No for Qwen3.8-Max, and a provisional yes for Qwen3.8-27B. The 2.4-trillion-parameter flagship whose weights arrive next week is datacenter hardware by any honest measure. The single sentence Alibaba attached about a 27B going open-weights alongside it is the part of this release that actually reaches a machine you own.

The August 2 announcement is the sequel to the July 19 preview tease we examined in our breakdown of the Qwen3.8 open-weights promise, which ended with a checklist of the artifacts that would turn Alibaba's claims into facts. Two weeks later, some of those artifacts are landing. This piece scores that checklist, runs the hardware math on both models, and lists exactly what to verify before download day.

What Alibaba Announced on August 2, and Where the July Checklist Stands

The announcement converts July's open-ended promise into a committed window: Qwen3.8-Max's open weights ship next week, per the official Qwen announcement and the accompanying Qwen blog post. The line that matters most for home hardware is delivered almost as an aside: "Qwen3.8-27B is also going open-weights to meet you all." That clause is doing more work for readers of this site than the 2.4T headline above it, and we will come back to it.

Alibaba also published the first standalone per-token pricing for the model: $2.00 per million input tokens, $6.00 per million output tokens, and $0.25 per million for implicit caching, per the August 2 announcement. During the preview era there was no per-token rate at all; access was bundled through subscription tiers on Alibaba's own platforms. A published API price is a small thing, but it is one of the concrete artifacts the July piece said to watch for. Reporting from the South China Morning Post adds that the model is now widely accessible through Alibaba Cloud's Model Studio APIs with a context window of up to 1 million tokens, and frames the release as Alibaba's return to open-sourcing its top-tier models after two closed flagships.

Here is the honest scorecard against the four artifacts our July analysis said would turn claims into facts. The release date moved from nothing to a committed week: landed, pending the actual drop. The API pricing moved from bundled subscriptions to published per-token rates: landed. The license text: still unpublished for both models. The active-parameter count and architecture details: still undisclosed. The model card and weight files on Hugging Face: still absent as of publication, and Qwen's Hugging Face organization remains the page to watch. Independent benchmarks: still none. Alibaba did publish a new benchmark chart with the announcement, and it deserves a careful read rather than a headline. On Alibaba's own slide, the model leads on research reproduction and cowork-style suites, and it also trails Claude Fable 5 by roughly 12 points on both of the hardest software-engineering benchmarks the chart includes. Every number on it is vendor-run. When a company's own marketing chart concedes double-digit gaps to a competitor on the hardest tests, believe the concession before the wins.

The 2.4T Math: Why the Max Is Not a Home Model

At 4-bit precision, 2.4 trillion parameters is roughly 1.2 terabytes of weights before context cache and runtime overhead, and no single consumer machine holds half of that. The table below puts the Max next to the only two open-weight models in its size class, with each cell limited to what is actually on the record.

Qwen3.8-Max figures are Alibaba's announcement claims as of August 2, 2026. Kimi K3 figures are from Moonshot's official model card. DeepSeek V4-Pro figures are from its April 2026 release documentation.

The precedent for what happens next was set two weeks ago. Kimi K3 ships natively at 4-bit, roughly 1.5TB, and the smallest calibrated build anyone has produced is Unsloth's dynamic 1-bit quantization at roughly 600GB with a 650GB-plus floor of combined RAM and VRAM. Apply the same roughly 60 percent reduction to a 1.2TB Qwen3.8-Max and you land near a 500GB-class 1-bit build with a memory floor still around half a terabyte. That is arithmetic, not a published build, and it is the optimistic case. Even then, the largest single-machine memory pool a consumer can buy, a 512GB Mac Studio, sits at the floor before the operating system takes its share. We worked through this exact wall in our Kimi K3 hardware reality check, and the conclusion transfers with the dial turned down about 14 percent: two-machine clusters, retired servers with 768GB of RAM, or datacenter accelerators.

One clarification the undisclosed architecture does not change: the missing active-parameter count decides speed and serving cost, not the memory floor. In a mixture-of-experts model, every expert has to sit in memory whether or not it fires on a given token. Sparse or dense, 2.4 trillion parameters occupy the same shelf space.

Qwen3.8-27B Is the Actual News for Your Hardware

One clause in the announcement reaches hardware people own, and it is the 27B. Alibaba's 27B line is not a random size. Qwen3.6-27B, released April 22, 2026, is the current dense 27B and the base much of the fine-tuning community standardized on; it is the safe-pick dense model in our guide to the best local AI models by VRAM tier. A successor in that slot, trained in the same family as a 2.4T flagship, is exactly the teacher-to-student pipeline that has carried frontier capability down to consumer hardware every cycle so far.

The sizing math is friendly. A 27B model quantized to 4-bit is roughly 14 to 16GB of weights, which fits a 24GB GPU with room for context, or a 32GB Mac or mini PC with headroom for the OS and a long session. That is the class of machine we specify tier by tier in our mini PC guide for local AI, and it means the interesting half of next week's drop is a lunch-break download, not an infrastructure project.

Now the unknowns, because they are real. Alibaba has not published whether Qwen3.8-27B is dense or a small mixture-of-experts, whether it inherits the Max's multimodal stack, what its exact file sizes are, or what license it ships under. The pattern is encouraging: recent open Qwen releases at 35B and below, including Qwen3.6-27B, shipped under Apache 2.0, one of the most permissive licenses in the industry. But a pattern is not a promise, and the K3 launch just demonstrated why: Moonshot's weights arrived under a bespoke license that one widely shared deployment guide misidentified as Apache 2.0. Read the LICENSE file in the repository before building anything on either model.

Check the Runtime Before You Download Anything

Weights on Hugging Face do not mean your software can load them, and the K3 launch demonstrated the gap in three ways inside one week. Moonshot's model card named vLLM, SGLang, and TokenSpeed as its serving engines and did not name llama.cpp, Ollama, or LM Studio. llama.cpp support arrived through a dedicated development branch rather than a stable release, which is why Ollama and LM Studio still cannot load the files. And the Kimi K3 listing that appeared in Ollama's library turned out to be a cloud service with a paid tier, a hosted API wearing a familiar command, not local inference.

So here is the checklist for Qwen3.8 drop day, in order. First, confirm mainline llama.cpp support is merged, not sitting in a pull request. Second, if you use Ollama, confirm the model tag has no cloud suffix before assuming your prompts stay on your machine. Third, open the LICENSE file and read it before anything commercial touches the weights. There is one reason for mild optimism on the 27B specifically: if it is a conventional dense Qwen, runtime support could be near-immediate, because the architecture is already supported everywhere. If it shares new architecture with the Max, expect a lag measured in weeks. Verify rather than assume.

The Download vs. Your Data Cap

A 1.2-terabyte flagship download is an entire month of a typical 1.2TB Xfinity data cap in one file, and nearly all of a 1.25TB Cox cap. Even a hypothetical 500GB-class 1-bit build would consume about 40 percent of a capped month before you open a browser. The 27B is the opposite story: 14 to 16GB is a rounding error on any cap and minutes on a gigabit line. We ran this exact math tier by tier, provider by provider, in our K3 download-time breakdown; every table in it applies here with smaller numbers. If your ISP meters you, know where you stand in the billing cycle before the repository goes live.

Frequently Asked Questions

Can you run Qwen 3.8 locally?

Not the Max. At 2.4 trillion parameters, Qwen3.8-Max needs roughly 1.2TB of memory at 4-bit precision, far beyond any consumer machine. Qwen3.8-27B is the realistic local target: a 27B quantized to 4-bit fits a 24GB GPU or a 32GB Mac or mini PC, pending the actual file sizes and runtime support when the weights land next week.

How big is the Qwen3.8-Max download?

No official sizes exist yet. The arithmetic: 2.4 trillion parameters is roughly 1.2TB at 4-bit precision and several times that at full precision, before any calibrated low-bit builds appear. For comparison, Kimi K3 at 2.8T shipped at roughly 1.5TB native, and its smallest third-party 1-bit build landed near 600GB. Exact Qwen figures arrive with the repository.

What is Qwen3.8-27B?

Qwen3.8-27B is the smaller open-weights model Alibaba announced alongside Qwen3.8-Max on August 2, 2026. Its architecture, file sizes, and license have not been published. It succeeds Qwen3.6-27B, the dense 27B released in April 2026 that became a standard base for the fine-tuning community, which is why a new 27B in this family is credible consumer-hardware news.

Is Qwen 3.8 open source?

Not yet, and open-weight is the more accurate term even after release: downloadable parameters, not training data or full training code. No license has been published for either Qwen3.8 model as of August 2, 2026. Recent open Qwen releases at 35B and below shipped under Apache 2.0, but treat that as a pattern to verify against the LICENSE file, not a promise.

Which is bigger, Qwen3.8-Max or Kimi K3?

Kimi K3, at 2.8 trillion total parameters against Qwen3.8-Max's vendor-claimed 2.4 trillion. K3's weights have been downloadable since July 27, 2026, which makes it the largest open-weight model available today. If Alibaba ships next week as committed, Qwen3.8-Max becomes the second-largest open release ever, well ahead of DeepSeek V4-Pro at 1.6 trillion.

What can I run at home while I wait?

Plenty. Open models in the 8GB-to-128GB range handle coding, document work, and light agent tasks well in 2026, and they keep improving on hardware you already own. Our guide to the best local AI models by VRAM tier matches a model to your memory, and our mini PC guide covers a dedicated, network-isolated box for the job.