Qwen 3.8 27B: How to Run This Open Model Locally

What is Qwen 3.8 27B?

Qwen 3.8 27B is a 27 billion parameter open-weight language model from Alibaba, released under an Apache 2.0 license on Hugging Face. Unlike the much larger Qwen 3.8 2.4T (a 2.4 trillion parameter model that’s effectively too big for any local machine), the 27B variant is sized to run on consumer hardware, specifically GPUs with 24 to 32GB of VRAM, or Apple Silicon Macs with enough unified memory. What makes it notable isn’t just that it’s runnable locally. It’s that benchmark results place it close to some proprietary frontier models, which is a rare combination for a model this size.

TL;DR

Qwen 3.8 27B is an open-weight, Apache 2.0 licensed model from Alibaba, distinct from the far larger 2.4T parameter Qwen 3.8 release that came out the week before it.
On the artificial analysis agentic index, Qwen 3.8 27B scored above GPT-5.6 Terra and Luna and only slightly behind Kimi K3, despite being small enough to run on a single high-end consumer GPU.
On the broader intelligence index (a composite of multiple benchmarks), it scored 52, landing just below GPT-5.6 Luna Max, DeepSeek V4, and Gemini 3.7 Flash.
Running it well requires 24 to 32GB of VRAM, meaning cards like the RTX 3090, 4090, or 5090, or a Mac Studio with sufficient unified memory; a 16GB card can technically load it but runs very slowly.
LM Studio is the simplest way to run it locally, offering multiple quantization options (4-bit, 5-bit, 6-bit, 8-bit) and a hardware check that tells you whether full GPU offload is possible before you download anything.
Quantization trades a small amount of model quality for a smaller file size and lower VRAM needs, so the general rule is to run the highest-bit version your hardware can fully offload.
An uncensored community variant is also available through LM Studio’s model search, and it behaves exactly as advertised: it will answer prompts that most mainstream models refuse.

REMY IS NOT

✕a coding agent
✕no-code
✕vibe coding
✕a faster Cursor

IT IS

✓a general contractor for software

The one that tells the coding agents what to build.

How does Qwen 3.8 27B compare to closed frontier models?

Benchmark placement is where this release gets interesting. On the artificial analysis agentic index, a benchmark that measures how well a model performs on agent-style tasks, Qwen 3.8 27B ranked above GPT-5.6 Terra and Luna, and only slightly behind Kimi K3, another open model but one large enough that it isn’t practical to run outside a data center.

On the intelligence index, which blends results from a wide set of benchmarks into a single composite score, Qwen 3.8 27B posted a 52. That puts it just under GPT-5.6 Luna Max, DeepSeek V4, and Gemini 3.7 Flash. The gap to those frontier-tier models is small. And crucially, every model that outranks Qwen 3.8 27B on that list is too large to run on a consumer GPU. That’s the real headline: it’s not the smartest model available, but it may be the smartest model you can actually load onto your own machine.

What hardware do you need to run it?

The practical requirement is 24 to 32GB of VRAM for a smooth experience. On Nvidia hardware, that effectively narrows the field to GPUs ending in “90”: the RTX 3090, 4090, and 5090, since these are the consumer cards with enough VRAM headroom. Apple Silicon Macs with large unified memory pools, such as the Mac Studio line, are also capable of running it.

It is technically possible to load the model on a 16GB card, but performance drops sharply, to the point of being impractical for regular use. If your GPU doesn’t have at least 24GB of VRAM, you’re better off choosing a more aggressively quantized version or a smaller model entirely.

How do you install Qwen 3.8 27B with LM Studio?

LM Studio is a free application built for running language models locally without needing to configure inference servers or manage Python environments by hand. The setup process is straightforward:

Download and install LM Studio.
Open the model search inside the app and search for Qwen 3.8 27B.
Select the official release from Qwen’s Hugging Face page rather than a third-party repack.
Choose a quantization level: 4-bit, 5-bit, 6-bit, or 8-bit are typically offered, along with the full unquantized 16-bit version if your hardware supports it.
Before downloading, LM Studio flags whether “full GPU offload” is possible on your specific machine, so you can check compatibility ahead of time rather than guessing.
Once downloaded, go to “My Models,” select the version you installed, and choose “Use in new chat” to start interacting with it.

The model supports the features people now expect from modern local LLMs, including visible reasoning/thinking steps and code generation, all running entirely on your own hardware with no API calls or subscription required.

What does quantization actually change?

Quantization compresses a model’s weights to reduce file size and memory requirements, at the cost of some precision. A model at 4-bit quantization takes up much less VRAM than the same model at 8-bit or 16-bit, but it may lose some accuracy or nuance in its outputs. The general guidance is to run the highest-bit (least compressed) version your GPU can fully offload to VRAM, since that preserves more of the model’s original capability. LM Studio’s built-in compatibility check removes most of the guesswork here by telling you upfront which quantization levels your system can realistically handle.

Is the uncensored variant worth using?

LM Studio’s model search also surfaces a community-built uncensored variant of Qwen 3.8 27B. This version strips out the guardrails that typically cause mainstream models to refuse certain requests. Testing an 8-bit uncensored build confirmed that it will respond to prompts that most consumer-facing models decline to answer, including requests other models are specifically trained to refuse. That flexibility comes with an obvious tradeoff: there’s no built-in safety layer, so anyone running it should understand they’re responsible for how they use it. It’s a good option for developers who want an unrestricted local sandbox, but not something to deploy in a consumer-facing product without adding your own moderation layer on top.

Is Qwen 3.8 27B worth running locally?

For anyone with a 24GB+ VRAM GPU or a high-memory Mac, yes. It offers benchmark performance close to several proprietary frontier models while running entirely offline, with no per-token cost and no data leaving your machine. The tradeoffs are real: hardware requirements exclude most laptops and budget GPUs, and quantized versions sacrifice some capability for accessibility. But for developers who want a capable, open-weight model that doesn’t require a data center, Qwen 3.8 27B represents one of the more compelling local options currently available.

Frequently Asked Questions

What’s the difference between Qwen 3.8 27B and Qwen 3.8 2.4T?

Qwen 3.8 2.4T is a 2.4 trillion parameter open-weight model, far too large for local hardware. Qwen 3.8 27B is a separate, much smaller 27 billion parameter release designed to run on consumer GPUs and high-memory Macs.

How much VRAM do I need to run Qwen 3.8 27B?

Around 24 to 32GB of VRAM for a good experience, which means GPUs like the RTX 3090, 4090, or 5090. It can load on 16GB cards but runs very slowly.

What is the easiest way to run Qwen 3.8 27B locally?

LM Studio is the simplest route. It’s a free app that lets you search for the model directly, choose a quantization level, checks whether your hardware supports full GPU offload, and lets you chat with the model once downloaded.

Does Qwen 3.8 27B outperform GPT and Gemini models?

On the artificial analysis agentic index, it scored above GPT-5.6 Terra and Luna. On the broader intelligence index it scored 52, just below GPT-5.6 Luna Max, DeepSeek V4, and Gemini 3.7 Flash, but ahead of many other models.

Is there an uncensored version of Qwen 3.8 27B?

Yes, a community-built uncensored variant is available through LM Studio’s model search. It removes standard content guardrails and will respond to prompts most mainstream models refuse.