How to Run Qwen3.8 27B Locally: Hardware, VRAM & Setup Guide (2026)

Qwen3.8 27B is Alibaba’s newest open-weight model, released in August 2026. It’s one of the strongest models you can realistically run on your own computer. The short version: you need about 18GB for the model at 4-bit, plus room for context, so a 24GB graphics card or a Mac with 32GB or more of unified memory. This guide covers exact memory needs by quantization and context length, which hardware runs it well, how to set it up with Ollama, LM Studio or vLLM, and when a cloud API makes more sense.

Key Takeaways

• Model: dense 27.8B parameters, 256K native context (extendable towards 1M), text, image and video input, Apache 2.0 licence.
• Memory: the Ollama download is ~18GB at Q4_K_M, 30GB at Q8_0 and 56GB at full BF16. Budget about 21–22GB total for Q4 with a 32K context.
• Best hardware: a 24GB GPU (used RTX 3090 or RTX 4090), an RTX 5090, or a Mac with 48GB+. A 32GB Mac works with a shorter context.
• Doesn’t fit fully: 16GB GPUs. Part of the model spills into system RAM, which is much slower.
• Thinking mode is on by default. Turn it off for faster, direct answers.

What Is Qwen3.8 27B?

According to Alibaba’s model card, Qwen3.8 is “the most capable generation in the Qwen open-model family to date”. The 27B version is a compact dense model: every parameter is used for every token, unlike the Mixture-of-Experts Qwen3.6 35B-A3B. It builds on the Qwen3.5 architecture with a hybrid attention design. Of its 64 layers, 48 use a linear-attention mechanism (Gated DeltaNet), and only 16 use standard full attention. That’s a big deal for local users, because only those 16 layers need a KV cache that grows with your conversation (see the memory section below).

Other highlights from the official model card:

• Native vision-language model: it understands images, documents, diagrams and even long videos.
• Thinking control: thinking mode is on by default and can be disabled per request. Reasoning depth can be tuned (xhigh, medium or low) on frameworks that support it.
• Context: 262,144 tokens natively, and Alibaba says it can be extended up to 1M tokens with YaRN on vLLM, SGLang or TokenSpeed.

How Good Is It? (Alibaba-Reported Benchmarks)

These scores come from Alibaba’s model card, not our own testing. They show a clear jump over the previous 27B model:

The biggest gains are in agentic coding and computer use, which makes it an excellent local model for coding assistants and agents. See also our guide to running Qwen coding models in VS Code and Cursor.

Memory Requirements

Download sizes are from Ollama’s model library and include the vision encoder. Totals are estimates.

Why its long context is cheap

For the 16 full-attention layers, each token of context stores 16 layers × 4 KV heads × 256 dimensions × 2 (keys and values) × 2 bytes. That’s roughly 64KB per token at 16-bit precision. So a 32K context adds about 2GB, 128K about 8GB and the full 256K about 16GB. The linear-attention layers keep a small fixed-size state instead of a growing cache. A classic dense model of this size would need several times more memory for the same context. You can explore other context sizes in our VRAM & KV Cache Calculator, which now models this hybrid design.

Which Hardware Runs It Well?

Generation speed is mostly limited by memory bandwidth, because each new token requires reading all ~16.7GB of the language-model weights at 4-bit. Using our rule of thumb tokens/sec ≈ bandwidth × 0.65 ÷ 16.7GB:

These are estimates from published specs, not measurements. Ollama’s default qwen3.8:27b build includes multi-token prediction (MTP), which can speed up generation beyond these baseline figures. Gains vary by prompt. Compare specific machines in the Local Hardware Matrix or on the Qwen3.8 27B model page. For buying advice, see our budget GPU guide and best Mac for local LLMs.

Mac users: by default, macOS only lets the GPU use part of unified memory (about two-thirds on a 32GB Mac). Our Mac VRAM Optimizer shows how much you can safely unlock.

How to Run Qwen3.8 27B

Option 1: Ollama (easiest)

1. Install Ollama from ollama.com.
2. Run the model. It downloads ~18GB the first time:
   ollama run qwen3.8:27b
3. Want faster answers without the thinking phase? Start it with ollama run qwen3.8:27b --think=false, or type /set nothink in the chat.
4. Need more context for coding or documents? Raise Ollama’s context length, for example by setting the OLLAMA_CONTEXT_LENGTH environment variable to 32768, if you have the memory headroom.

On Apple Silicon, Ollama also offers MLX builds (qwen3.8:27b-mlx), which are often faster on Macs.

Option 2: LM Studio (graphical app)

Search for “Qwen3.8 27B” in LM Studio, choose a 4-bit GGUF (or an MLX build on a Mac), and it will show you which versions fit your hardware. See our comparison of local LLM runners.

Option 3: vLLM (GPU servers and teams)

For serving multiple users on NVIDIA or AMD GPUs, Alibaba recommends dedicated engines such as vLLM, SGLang or TokenSpeed. A minimal start is vllm serve Qwen/Qwen3.8-27B, which exposes an OpenAI-compatible API. Check the official vLLM recipe for recommended flags and the YaRN settings needed to go beyond 256K context.

Recommended Settings

From the official model card:

• Thinking mode: temperature 1.0, top_p 0.95, top_k 20, presence_penalty 0.0.
• Non-thinking mode: temperature 0.7, top_p 0.80, top_k 20, presence_penalty 1.5.
• If you see endless repetition, raising presence_penalty (0–2) can help, though high values may cause language mixing.

Not sure which 4-bit format to choose? Our quantization guide explains the trade-offs.

Local vs API: Which Is Cheaper?

If you mainly want the model’s quality rather than privacy, Groq serves Qwen3.8 27B at $0.80 per 1M input tokens and $4.00 per 1M output tokens, at a published ~450 tokens/sec (September 2026). A heavy solo developer using 15M input and 3M output tokens a month would pay about $24/month. A used-RTX-3090 PC that runs the model locally costs roughly $1,700 in 2026, so it would take years to pay back on price alone. Local makes sense when you need privacy, offline access or very high volume, or already own the hardware. Run your own numbers in the API Pricing & Break-Even Calculator.

FAQ

Can I run Qwen3.8 27B on a 16GB GPU?

Not fully. The 4-bit model alone is about 18GB, so part of it must run from system RAM, which is much slower. For 16GB cards, gpt-oss-20b or Gemma 4 12B are better fits. See How Much RAM and VRAM Do You Need?

Will it run on a Mac with 32GB?

Yes, at 4-bit with a modest context, but it’s tight because macOS reserves part of memory for the system. A 48GB or 64GB Mac is much more comfortable.

Does Qwen3.8 27B understand images?

Yes. It’s a native vision-language model, and Ollama lists text and image input for it. Reading documents locally with Ollama covers which vision models are built for that job.

How do I turn off thinking mode?

In Ollama, use --think=false or /set nothink. Through an OpenAI-compatible server, pass enable_thinking: false in the chat template settings, as shown on the model card.

Is it free for commercial use?

Yes. Qwen3.8-27B is released under the Apache 2.0 licence.

Sources

Qwen3.8-27B model card (Hugging Face), Ollama: qwen3.8, Ollama docs: thinking, Groq models and pricing. Hardware bandwidth figures are from Apple, NVIDIA and AMD specifications. Benchmarks are Alibaba-reported. Speed figures are BFWR estimates, not measurements.

Published September 11, 2026.