Best PC for Qwen 3.8 27B: The VRAM you really need
"It runs on 17 GB."
That's true. And it's misleading.
Qwen 3.8-27B loads on a 16 GB card. Its performance varies depending on how you use it.
Qwen 3.8-27B is currently the most discussed open model. A dense model with 27 billion parameters, published by Alibaba under an Apache 2.0 license, it combines reasoning, agentic capabilities, coding, and vision, with a native context window of 256,000 tokens.
Everywhere, the same figure is cited: it runs on approximately 17 GB. This figure is accurate. It simply doesn't tell the whole truth.
The 17 GB trap
Weights are the floor, not the total
The advertised 17 GB corresponds to the weight of the 4-bit quantized model. This is what is loaded at startup, before you've even asked a single question.
Then comes the KV cache. It stores the conversation context and grows with each token processed. In a long context, it alone can double the memory footprint.
In addition, there are execution buffers, CUDA graphs, the visual encoder if you're processing images, and the memory consumed by each concurrent session. Multi-token prediction, if enabled, requires an additional 1 to 2 GB of margin.
Hence the difference between "it loads on my card" and "it handles my workload".
Illustrative proportions. Actual cache footprint depends on context length, number of concurrent sessions, and cache quantization options.
VRAM actually needed
| Format | Weights only | Minimum card | Comfortable card |
|---|---|---|---|
| 4-bit (GGUF, NVFP4) | approx. 16 to 19 GB | 16 GB, short context | 24 to 32 GB |
| 4-bit + long context | 17 GB + KV cache | 24 GB | 32 GB |
| FP8 | approx. 27 to 28 GB | 32 GB | 48 GB and up |
| BF16 (native precision) | approx. 54 to 56 GB | 2 cards or 96 GB | 96 GB ECC |
Three scenarios, three machines
Discovery and conversational use
16 GBYou want to test Qwen 3.8-27B, use it for writing, summarizing, asking questions, processing reasonably sized documents.
A 16 GB card is sufficient in 4-bit, provided you stick to moderate contexts. This is the most economical entry point for serious individual use.
The limit is felt as soon as you load very long documents or engage in long conversations.
Code agent and tools
32 GBThis is where Qwen 3.8-27B excels: agentic coding, tool calls, working on a codebase.
An agent re-reads its context at each step. Ten rounds of tool calls on a large repository will swell the cache very quickly. On 16 GB, the session eventually saturates.
With 32 GB, you maintain margin for extended context and good throughput. This is the reference configuration for daily professional use.
Massive context, vision, multi-user
96 GB and upLeveraging 256K context tokens, processing images, or serving multiple people concurrently is a different category.
The KV cache then becomes the dominant item, far ahead of the weights. A unified 128 GB memory or a professional 96 GB ECC card become relevant, as does FP8 to preserve quality.
This is also the threshold from which fine-tuning the model on your own data can be considered.
Our recommendation
Radiance CoreAI 64 — RTX 5090 32 GB
- GPU RTX 5090, 32 GB GDDR7
- Bandwidth 1,792 GB/s
- CPU Ryzen 9 9950X3D
- RAM DDR5 64 GB
- Storage NVMe 1 TB
- Power Supply 1,200 W 80+ Gold
32 GB: the 4-bit model, extended context, and margin for agents.
Running it in practice
The fastest way is through Ollama. The download is approximately 18 GB.
# Direct launch ollama run qwen3.8:27b # Variant with extended context and quantized cache cat > Modelfile <<'EOF' FROM qwen3.8:27b PARAMETER num_ctx 65536 EOF ollama create qwen3.8-long -f Modelfile
Qwen 3.8-27B is a hybrid model: it has a reasoning mode and a direct instruction mode, with distinct recommended settings.
| Mode | temperature | top_p | top_k | Usage |
|---|---|---|---|---|
| Reasoning | 1.0 | 0.95 | 20 | Complex problems, code, analysis |
| Direct instruction | 0.7 | 0.80 | 20 | Writing, quick answers |
Our entire range
Each machine is hand-assembled in Auriol (13390) and fully configurable, via the online configurator or by quote at contact@radiancesystems.eu.
Individual workstations
CoreAI 16 — RTX 5060 Ti 16 GBQwen 3.8-27B in 4-bit, short context. The entry point.
€1,703
CoreAI 32 — RTX 5070 Ti 16 GBSame VRAM envelope, more CPU and RAM for the pipeline.
€2,442
CoreAI 64 — RTX 5090 32 GBThe benchmark: agents, extended context, vision, maximum throughput.
€6,042
Compact server
NVIDIA GB10 AI Mini Server128 GB unified: 256K context without constraints, multiple users.
€3,999
Rack and multi-GPU
CoreAI Rack — 2 × RTX 5090 (64 GB)Comfortable FP8, multiple concurrent sessions, 4U rack format.
€11,221
Pro AI Ultra — Threadripper PRO96 GB ECC: native precision BF16, model fine-tuning.
€20,213
CoreAI 128 Rack — 2 × RTX 6000 (192 GB ECC)Production, numerous users, training, ECC reliability.
€27,980
In brief
How much VRAM for Qwen 3.8-27B?
16 GB for 4-bit short context. 32 GB for agents and extended context. 96 GB and up for native precision or fine-tuning.
Why isn't 16 GB always enough?
Because the advertised 17 GB only accounts for the weights. The KV cache, execution buffers, and visual encoder add up, and the cache grows with context length.
Is Qwen 3.8-27B free?
Yes, Apache 2.0 license, commercial use included. You only pay for the hardware.
Does it process images?
Yes, it includes a visual encoder. Vision consumes additional memory, which should be factored into sizing.
Can it be fine-tuned on custom data?
Yes, this is one of the advantages of a dense model of this size. However, expect significantly more VRAM than for inference: configurations with 96 GB ECC and up are most suitable.
The memory values indicated are estimates derived from ecosystem feedback and vary depending on the inference engine, quantization method, context length, and number of concurrent sessions. We are happy to advise you based on your actual usage.
