Qwen 3.8 27B: Specs, Hardware Requirements, and How to Run It (2026)

The 2.4 trillion parameter Qwen 3.8-Max got the headlines. The model most teams will actually run is the other one Alibaba announced alongside it: Qwen3.8-27B, going open-weight within about a week of the August 3 launch.

Nobody self-hosts a 2.4T model. A 27B is a different story: it fits on a single rented GPU, it fine-tunes on realistic budgets, and its predecessor is one of the most-praised local models of the year. Here is what's confirmed about Qwen3.8-27B so far, the hardware math to plan around, and how to be ready to serve it the day the weights land.

TL;DR

• Announced: August 3, 2026, alongside Qwen 3.8-Max. Alibaba committed to releasing open weights for both within about a week
• Not out yet: as of this writing there is no Hugging Face repository, no license, and no published spec sheet
• Specs: unconfirmed. Alibaba has not said whether it's dense like Qwen3.6-27B or sparse MoE, or what the context window is
• Early third-party signal: Unsloth says quantized builds will run locally on roughly 17GB of RAM or VRAM, which is consistent with a 4-bit 27B-class model
• Hardware planning: a 27B-class dense model needs roughly 54GB of VRAM at BF16, ~27GB at FP8, and ~14 to 16GB at 4-bit, before KV cache
• The contrast that matters: Kimi K3 self-hosting starts at 1.56 TB of weights and cluster scale. This is the open frontier-lab release you can run on one GPU
• The flagship alternative: Qwen 3.8-Max is API-only at $2 in / $6 out per million tokens. Full breakdown in our Qwen 3.8 API access guide

What Qwen3.8-27B Is, and What's Actually Confirmed

Confirmed: it exists, and it's going open-weight. Alibaba announced Qwen3.8-27B at the August 3 Qwen 3.8-Max launch and committed to releasing weights for both models within about a week. That commitment, not the flagship's parameter count, is the news for anyone who runs their own inference.

Also confirmed, by lineage: what it's replacing. Qwen3.6-27B, the current open 27B, is a dense model that became a community favorite for local coding and agent work, the kind of model people describe as punching above its weight class. Qwen3.8-27B arrives as its successor, presumably carrying the 3.8 generation's training improvements down to self-hostable scale.

Not confirmed, and worth being honest about: everything else. Alibaba has not published architecture details (dense or MoE), context length, benchmark scores, or a license. Nothing is on Hugging Face yet. The one early third-party signal comes from Unsloth, who say quantized builds will run on roughly 17GB of RAM or VRAM. That's consistent with a 4-bit quant of a 27B-class model, and Unsloth typically ships GGUF quants within hours of a Qwen release, but treat it as a preview claim until the weights exist.

We track the full Qwen 3.8 launch picture, Max included, in our release breakdown and benchmarks tracker.

Qwen 3.8 27B Hardware Requirements

Until the spec sheet drops, plan around a 27B-class dense model, the same shape as Qwen3.6-27B. The memory math:

Two caveats that decide whether these numbers hold in practice.

First, weights are the floor, not the total. KV cache comes on top and scales with context length and concurrency. A 27B model serving long-context requests to multiple users can double its memory footprint. If Qwen3.8-27B inherits a long context window, the gap between "loads on my GPU" and "serves my workload on my GPU" will be significant. Our breakdown of what actually limits inference speed covers the mechanics.

Second, if the model ships as sparse MoE instead of dense, the memory math stays roughly the same and the compute math gets better. MoE saves you compute per token, not memory: all parameters live in VRAM regardless of how few activate. We walked through the same principle at the opposite end of the scale in our Kimi K3 hardware requirements breakdown, where 2.8T parameters need 1.56 TB of memory despite only 104B activating per token.

That K3 comparison is the point of this release. K3 self-hosting starts at 64-plus accelerators and cluster-scale orchestration. A 27B runs on one rented GPU. If Alibaba ships both drops as promised, the Qwen family will cover both ends of the self-hosting spectrum while K3 remains cluster-only.

How to Run It on Day One

The serving playbook for a 27B-class Qwen is well established, because the ecosystem has been running its predecessor for months.

vLLM or SGLang for production serving. Qwen releases have shipped with day-one support in both engines historically, and there's no reason to expect 3.8 to break the pattern. If you're choosing between them, our vLLM vs SGLang comparison covers the decision; either gives you an OpenAI-compatible endpoint out of the box, so the model slots into existing client code. Deployment mechanics are in our vLLM production Docker guide.

Quantized local builds for evaluation. Expect community GGUF and AWQ quants fast, likely from Unsloth first. That's the 17GB path: fine for testing quality on your workload, not what you want under production traffic.

The GPU decision. For serious serving, FP8 on a 48GB card is the likely sweet spot for a 27B, with BF16 on an 80GB card when quality headroom matters. Our best GPUs for LLM inference guide compares the options; the same sizing logic we used for running Qwen 3.7 in production applies here at smaller scale.

If you want the hardware ready before the weights drop, single GPUs in every class above are available on Yotta GPU Pods by the hour, which is the cheapest way to find out what a new model actually needs before committing to anything.

Qwen3.8-27B vs Qwen 3.8-Max

Same generation, opposite deployment stories.

Qwen 3.8-Max is API-only: 2.4T parameters, ~95B active, served by Alibaba at $2 per million input tokens and $6 per million output. You get the flagship's capability with zero infrastructure, and no control. Qwen3.8-27B will be the opposite: weights you hold, hardware you choose, fine-tuning rights, and capability that no 27B will match a 2.4T flagship on.

The practical split most teams land on: the big model through an API for the hardest reasoning, the small model self-hosted for volume work. Until 3.8-Max shows verified benchmarks, there's an extra reason to keep that split loose: nobody knows yet how much capability the API premium actually buys. Our Qwen 3.8 vs Kimi K3 comparison covers the evidence situation in full.

Frequently Asked Questions

When will Qwen 3.8 27B be released? Alibaba committed to releasing open weights within about a week of the August 3, 2026 launch announcement. As of this writing the weights have not landed on Hugging Face. This post updates the day they do.

Can I download Qwen 3.8 27B from Hugging Face? Not yet. No repository, license, or model card has been published. When it lands, expect quantized community builds to follow within hours.

What GPU do I need to run Qwen 3.8 27B? Planning math for a 27B-class model: ~54GB VRAM at BF16 (80GB card), ~27GB at FP8 (48GB card), ~14-16GB at 4-bit (24GB card), plus KV cache on top. Official specs may shift these numbers.

Will Qwen 3.8 27B run on an RTX 4090? Quantized 4-bit builds should, based on the ~17GB figure Unsloth has previewed. BF16 will not fit on a 24GB card.

Is Qwen 3.8 27B dense or MoE? Not disclosed. Its predecessor Qwen3.6-27B is dense. Either way the VRAM requirement is set by total parameters, not active ones.

What license will Qwen 3.8 27B use? Unknown. Alibaba has not named a license for either 3.8 open-weight release. Prior open Qwen releases have used permissive licensing, but don't build a compliance plan on a pattern.

Is Qwen 3.8 27B better than Qwen3.6-27B? No benchmarks exist yet for either 3.8 model. The reasonable expectation is generational improvement at the same size; the honest answer is wait for scores or run your own eval.

Bottom Line

Qwen3.8-27B is the practical half of the Qwen 3.8 launch: the model you'll actually be able to download, fine-tune, and serve on a single GPU, from the generation Alibaba claims is second only to Claude Fable 5. What's missing is everything that lets you verify the claim: weights, specs, license, benchmarks. All of it is promised within days.

The preparation that costs nothing: know your precision target, know your GPU, and have your serving stack ready. When the weights land, the teams that planned the memory math run the model that afternoon. This post gets updated the day it happens.