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, now open-weight on Hugging Face as of August 13-14.

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 serve it well now that the weights are out.

TL;DR

• Announced: August 3, 2026, alongside Qwen 3.8-Max. Alibaba committed to releasing open weights for both within about a week
• Update: the weights are out. Released August 13-14 on Hugging Face under Apache 2.0 at the official Qwen/Qwen3.8-27B repo, with a surprise vision encoder and 262k native context. Our local setup guide covers running it today.
• Confirmed by the ecosystem: quantized builds run locally on roughly 17GB of RAM or VRAM; Ollama's build is an 18GB download
• Hardware planning: a 27B-class dense model needs roughly 56GB of VRAM at BF16, ~28GB 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. For the other end of the market, GPT-6 Astra launched at $10 in / $50 out, 50x the input price of OpenAI's own budget tier.

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

Confirmed: it exists, and the weights are out. 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.

The spec sheet arrived with the weights, and it over-delivered: 27B parameters, dense with hybrid attention, a vision encoder (the release is multimodal, which nobody promised), 262k native context extensible to 1M, and an Apache 2.0 license. The official repo is Qwen/Qwen3.8-27B, and the model card publishes benchmark results, including 61.7 on SWE-bench Pro.

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

Qwen 3.8 27B Hardware Requirements

The spec sheet is out: 27B parameters, dense. 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.

The release settled the architecture question: dense, so the full 27B sits in VRAM and the table above is the real math. The MoE memory principle still matters 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. With both drops shipped, the Qwen family now covers 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. Unsloth's quants landed within a day, and their fine-tuning support followed. Our Unsloth fine-tuning guide covers that path. That's the 17GB path: fine for testing quality on your workload, not what you want under production traffic. We put together the step-by-step local setup, Ollama, GGUF, and Mac paths included, so you can be running it today.

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.

The complete deployment picture, API versus self-hosted and the cost crossover, is in the full production guide.

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 in practice: 2.4T parameters, ~95B active, served by Alibaba at $2 per million input tokens and $6 per million output, and now also live on Yotta AI Gateway. The 27B is on the Gateway too if you want the API route without self-hosting. You get the flagship's capability with zero infrastructure, and no control. Qwen3.8-27B is 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. For the complete side-by-side, including where Qwen 3.7-Max fits while 3.8 gets verified, see the full lineup comparison.

Frequently Asked Questions

When will Qwen 3.8 27B be released? It's out. The weights landed on Hugging Face August 13-14, 2026, under Apache 2.0, about ten days after the August 3 announcement.

Can I download Qwen 3.8 27B from Hugging Face? Yes: the official repo is Qwen/Qwen3.8-27B. Community quantized builds are already listed across llama.cpp, LM Studio, Jan, and Ollama. Our local setup guide covers the fastest path to running it.

What GPU do I need to run Qwen 3.8 27B? Confirmed math for the 27B release: ~56GB VRAM at BF16, ~28GB at FP8, ~14-16GB at 4-bit, plus KV cache on top.

Will Qwen 3.8 27B run on an RTX 4090? Yes: quantized 4-bit builds run on a 24GB card, and the Ollama build is an 18GB download. BF16 will not fit on a 24GB card.

Is Qwen 3.8 27B dense or MoE? Dense, 27B parameters with hybrid attention, plus a vision encoder.

What license will Qwen 3.8 27B use? Apache 2.0, confirmed at release. Commercial use, modification, and redistribution are all permitted.

Is Qwen 3.8 27B better than Qwen3.6-27B? The model card publishes benchmark results now, including 61.7 on SWE-bench Pro. Independent replication is pending, so run your own eval before switching production workloads, but there's finally data to compare.

How does Qwen 3.8 27B compare to DeepSeek V4 Flash?
On Artificial Analysis's Intelligence Index they score the same, 52 each, but the 27B runs on one GPU while DeepSeek V4 Flash needs two H200s, and the 27B takes image and video input where DeepSeek is text only. DeepSeek is the faster of the two on hosted APIs. The full head-to-head, including Qwen 3.8-Flash-Next, is in Qwen 3.8 vs DeepSeek V4 Flash.

Bottom Line

Qwen3.8-27B is the practical half of the Qwen 3.8 launch, and it delivered: Apache 2.0 weights, a vision encoder, 262k context, and published benchmark numbers, all downloadable today and all runnable on a single GPU. The teams that planned the memory math are running it this afternoon; the local setup guide gets you there.