VRAMGlass contains affiliate links. As an eBay Partner, we may be compensated if you make a purchase.As an eBay Partner, we may be compensated.Learn more

GPUs tracked57Price history49 daysPrice points (24h)43,303 pointsBiggest 24h dropGeForce RTX 3060 12GB-11.1%Biggest 24h riseCMP 170HX 40GB+19.8%Updated
$ USD

How to Deploy Qwen3.6 27B Locally

Qwen3.6 27B is a 27B dense model that Qwen released in April 2026, and it understands images. Qwen3.8 27B, released in August, keeps the same architecture with nearly the same size and speed but more capability, so new deployments should go straight to 3.8; 3.6 makes sense for workflows already tuned around it. On a 24GB card, its 4-bit version reaches about 117K of context.

Qwen3.6 27B details

  • Min VRAM at 4-bit19 GB · 8K
  • Parameters27.8B
  • Max context256K
Architecture
Dense
License
Apache-2.0
Launch date
Apr 22, 2026
Vendor
Alibaba
Hugging Face repo
Qwen/Qwen3.6-27B
GGUF repo
unsloth/Qwen3.6-27B-GGUF
NVFP4 repo (Blackwell)
nvidia/Qwen3.6-27B-NVFP4
MLX repo (Apple)
mlx-community/Qwen3.6-27B-4bit
Ollama tag
qwen3.6:27b
Layers
64
Hidden size
5,120
KV heads
4
Head dim
256

How much VRAM does Qwen3.6 27B need for local deployment?

QuantizationWeightsGB8K contextGB32K contextGB128K contextGB256K context (max)GB
1-bit7est.8.610.216.725.3
2-bit11.8GGUF13.41521.530.1
3-bit14.5GGUF16.117.724.232.8
4-bit16.8GGUF18.42026.535.1
5-bit19.5GGUF21.122.729.237.8
8-bit28.6GGUF30.231.838.346.9
BF16/FP1653.8GGUF55.45763.572.1

Full calculation rules are on the Methodology page.

Can a single GPU run Qwen3.6 27B?

A single 24GB card runs it at 4-bit. The RTX 3090, 4090 and RX 7900 XTX all fit it at about 40 tokens/s, with room for roughly 117K of context. The one shortfall is that Qwen recommends keeping at least 128K free for thinking, and Unsloth warns that too small a window can produce garbled output. To get there on a 24GB card, drop to 3-bit, or move to the 32GB RTX 5090, where 4-bit gets close to the full window. 16GB cards are limited to 2-bit, and 3.6's low-bit files are noticeably larger than 3.8's, so the fit is tighter. If you have a 16GB card and are starting fresh, Qwen3.8 27B is the better pick.

eBay
DevicePriceSpeed t/sMax context
Tesla V100 32GB32 GB$89518 listings38.3245K
Radeon RX 7900 XTX24 GB$1,25029 listings40.7117K
GeForce RTX 309024 GB$1,799107 listings39.7117K
GeForce RTX 409024 GB$3,75868 listings42.6117K
Ryzen AI Max+ 395 128GBWhole system · 128 GB$4,46420 listings11.3256K
RTX A600048 GB$4,89916 listings32.9256K
Mac Studio M5 Ultra 96GBWhole system · 96 GB$7,27412 listings51.2256K
MacBook Pro M5 Max 128GBWhole system · 128 GB$7,54920 listings26.6256K
DGX Spark 128GBWhole system · 128 GB$8,24022 listings12.1256K
GeForce RTX 509032 GB$8,395123 listings72.3245K
RTX PRO 600096 GB$22,50011 listings72.3256K

Only common GPUs and systems are listed, one set for models under 50B and another for larger ones, sorted by price from low to high.

What to know before running Qwen3.6 27B locally

  1. How it differs from 3.8 27B

    The two have identical architectures; the differences are all in training and usage. 3.8 is stronger at coding and long agent workflows, adds reasoning_effort to adjust thinking depth, and has “preserved thinking” on by default. 3.6 keeps only the most recent turn's thinking by default; to keep all of it, set preserve_thinking in chat_template_kwargs.

  2. Sampling parameters

    Same as 3.8 27B: temperature 1.0, top_p 0.95, top_k 20 and presence_penalty 0 in thinking mode, and 0.7, 0.8, 20 and 1.5 in non-thinking mode. 3.6 also has a separate set for precise coding work such as front-end development, with temperature lowered to 0.6.

  3. Framework versions and known issues

    Qwen recommends vLLM 0.19.0 or SGLang 0.5.10 or later. Unsloth warns against llama.cpp builds compiled with CUDA 13.2, which may output garbage; setting context too small can also produce garbage, and Qwen recommends leaving at least 128K for thinking. The Ollama tag is qwen3.6:27b.

  4. MTP speculative decoding

    llama.cpp now supports its MTP speculative decoding; just use Unsloth's MTP GGUF. Unsloth says generation gets 1.4 to 2.2x faster for only about 1 GB more. This site's speeds are estimated without MTP.

Launch commands

Ollama

  • Windows
  • macOS
  • Linux
Terminal
ollama run qwen3.6:27b

llama.cpp

  • NVIDIA
  • AMD
  • Mac
  • CPU
Terminal
llama-server -hf unsloth/Qwen3.6-27B-GGUF:Q4_K_M -c 8192

vLLM

  • Linux
  • NVIDIA
  • AMD
Original weights
Terminal
vllm serve Qwen/Qwen3.6-27B
NVFP4 · Blackwell

Blackwell GPUs (RTX 50 series, RTX PRO 6000, B200) run NVFP4 weights natively, at about 30% of the original size.

Terminal
vllm serve nvidia/Qwen3.6-27B-NVFP4

MLX

  • Apple Silicon

Vision models run on mlx-vlm: on a Mac, pip install mlx-vlm first; this command downloads the MLX weights from Hugging Face and starts a server.

Terminal
mlx_vlm.server --model mlx-community/Qwen3.6-27B-4bit

File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology

Other models in the Qwen family

ModelTotal paramsActive params4-bit size
Qwen3 8BDense8.2B8.2B5 GB
Qwen3.5 9BDense9.7B9.7B5.7 GB
Qwen3.8 27BDense27.8B27.8B16.5 GB
Qwen3 32BDense32.8B32.8B19.8 GB
Qwen3.6 35B A3BMoE36B3B22.1 GB
Qwen3.8 Flash NextMoE176B6B111.3 GB
Qwen3.8 2.4T A95BMoE2,446B95B1,310.9 GB

About Qwen