PC for Qwen 3.6: 27B dense or 35B-A3B MoE, which to choose?
Qwen 3.6 is the most talked-about open-source model right now. But before choosing a PC to run it, you need to answer one question: the 27B dense version, or the 35B-A3B MoE version? Everything else follows from that.
Both belong to the same family. Both fit on a consumer-grade card in 4-bit. Both are under the Apache 2.0 license. But they behave very differently in practice.
This guide answers that question first, then tells you the right PC for your choice.
Qwen 3.6: what you need to know
Qwen 3.6 is Alibaba's latest generation of open models. It is a multimodal model (text, image, video), with hybrid reasoning, and an exceptional native context window of up to 1 million tokens.
Two open variants are worth considering for local use:
Qwen3.6-27B
Dense — coding quality
- 27 billion parameters, all active
- Fits in about 16.8 GB in 4-bit
- Stable and predictable behavior
- Better at coding and following instructions
- Includes vision
- Benefits from DFlash acceleration (recent NVIDIA)
Qwen3.6-35B-A3B
MoE — speed
- 35 billion total, 3 billion active per token
- Computational cost close to a 3B model
- Quality close to a dense 35B model
- Very fast: over 100 tokens/s reported on high-end GPUs
- About 21 GB in 4-bit, requires 24 GB of VRAM
- Ideal for agents and fast tool-use
Dense or MoE: how to choose
VRAM: what you really need to plan for
The figures below are for a short to medium context. Warning: with extended context, the KV cache significantly increases the required memory.
| Variant | Quantization | VRAM (short context) | Recommended card |
|---|---|---|---|
| Qwen3.6-27B dense | Q4_K_M | about 16.8 GB | 16 GB tight, 24 GB comfortable |
| Qwen3.6-27B dense | Q5_K_M | about 20 GB | 24 GB |
| Qwen3.6-35B-A3B MoE | Q4_K_M | about 21 GB | 24 GB |
| Qwen3.6-35B-A3B MoE | Q5_K_M | about 26 GB | 32 GB |
| Either one, very long context | Q4 + quantized cache | +20 to 40 GB of KV cache | 32 GB and more |
Run Qwen 3.6 in two minutes
The easiest way is through Ollama. Choose the variant according to your card:
# 27B dense variant (16 GB and more) ollama run qwen3.6:27b # 35B-A3B MoE variant (24 GB and more) ollama run qwen3.6:35b-a3b # For agent or coding use, we often use llama.cpp # with quantized KV cache for long context: llama-server -m qwen3.6-35b-a3b-Q4_K_M.gguf \ --cache-type-k q8_0 --cache-type-v q8_0 \ --ctx-size 65536 --n-gpu-layers 99
What PC for Qwen 3.6
The choice of machine follows directly from the targeted variant. Here are our adapted stations, assembled in Auriol (13390) and delivered throughout the EU, with Ollama and Open WebUI preinstalled on request. For the heaviest variants of Qwen, our AI stations and servers go up to multi-GPU configurations.
CoreIA RTX 5060 TiQwen3.6-27B dense in Q4, at the memory limit. The entry point.
€2,068
CoreIA RTX 5090The benchmark: 35B-A3B in Q5, long context, over 100 tok/s.
€10,291
CoreIA GB10 (ASUS Ascent GX10)128 GB unified for 1M token context without constraints.
€6,999
Why run Qwen 3.6 locally
Beyond privacy, Qwen 3.6 locally offers concrete advantages for anyone who wants a serious AI on hand.
- No recurring cost. No subscription, no billing per token. Once the machine is acquired, use is unlimited.
- Private data. Your prompts, your code, your documents never leave your network.
- Top-tier quality. Qwen 3.6 rivals the best open models on coding, reasoning, and agent tasks.
- Massive context. Up to 1 million tokens natively, to process entire codebases or long documents.
- Apache 2.0 license. Free commercial use, without restrictions.
In short
27B dense or 35B-A3B MoE?
Dense for coding and reliable tool-use on 16 GB. MoE for maximum speed on 24 GB and more.
What minimum VRAM?
16 GB for the dense 27B. 24 GB for the 35B-A3B. 32 GB for long context or Q5.
Is Qwen 3.6 free?
Yes, open source under Apache 2.0. You only pay for the hardware.
Can it be used for coding and agents?
Yes, it is one of its strong points. For long agent loops, prefer the dense 27B, which is more consistent.
Do you need a powerful machine for a 1M token context?
Yes: the KV cache can add 20 to 40 GB. The GB10 mini-server with its 128 GB of unified memory is the best option for this requirement.
