Qwen3.8 27B VRAM requirements: can you run it locally?

Qwen3.8 27B is Alibaba Qwen's dense mid-size open-weight model. Its Hugging Face repo was created on August 5, 2026, it is Apache 2.0 licensed, and Hugging Face lists 27,781,427,952 parameters.^[1] Its config.json sets a maximum of 262,144 positions (256K tokens).^[3] It is one of the most-downloaded open models this month, with about 7.7 million Hugging Face downloads in 30 days.^[1]

Memory needed at each quantization

Quantization stores each weight in fewer bits. Smaller files fit on smaller GPUs, at some cost in quality, and the cost grows as you go below about 4 bits. File sizes below are the actual GGUF downloads on Hugging Face (1 GB = 1,000,000,000 bytes).^[4] The vision projector (mmproj, about 0.93 GB) is a separate file you only need for image input.^[4]

Context length adds memory (the KV cache)

Qwen3.8 27B is a hybrid model. Of its 64 layers, 48 use linear attention, which keeps a fixed-size state, and only every fourth layer (16 in total) uses full attention with a growing key/value cache.^[3] With 4 KV heads and a head size of 256, an FP16 cache takes 2 × 16 × 4 × 256 × 2 bytes = 65,536 bytes per token, far less than most dense models of this size:

The fixed linear-attention state is not included above; it does not grow with context. Even the full 256K context needs only about 17.2 GB of cache on top of the weights.

Which GPUs and Macs it fits on

Each cell compares the weights plus an 8K-token KV cache with the device's memory.^{[5][6][7][8][9][10][11]}

Column headers show weights + 8,192-token KV cache.

How to read it: "Fits" means weights + 8K cache use under 90% of memory. "Tight" means 90–100%: it may load, but the engine's own working memory or a longer context can push it over. "No" means it won't fit on the GPU alone. The 90% line is our own rule of thumb, not a vendor figure. On a Mac, macOS also needs some of the unified memory, so treat Mac rows as optimistic.

How to run it

With llama.cpp or LM Studio, download a GGUF file from the quant list above, for example Qwen3.8-27B-Q4_K_M.gguf (17.4 GB) for a 24 GB card or Qwen3.8-27B-Q3_K_M.gguf (13.4 GB) for a 16 GB card.^[4] Ollama lists the model under the qwen3.8 family.^[13] Hybrid linear-attention models need a recent engine build, so update before you load it.

If the model is bigger than your VRAM, these tools can keep some layers in system RAM, but every token then has to wait on the slower memory, so expect much lower speed.

Rent a GPU for this model

A 24 GB GPU is enough for Q4_K_M, so renting is cheap. On-demand Secure Cloud prices listed on Runpod's pricing page:^[12]

Prices checked on 2026-10-23. They change often, so confirm on the provider's page. A 48 GB card runs Q4_K_M tightly; 80 GB gives room for longer context or Q5/Q6.

Where to rent or buy

Frequently asked questions

Can I run Qwen3.8 27B on an RTX 3090 or RTX 4090?

Yes. The Q4_K_M weights are 17.4 GB and these cards have 24 GB, so with an 8K context it uses about 75% of the card.^[4][7][8] Q5_K_M (20.9 GB) also fits; Q6_K (23.9 GB) does not.

Can I run Qwen3.8 27B on a 16 GB or 12 GB GPU?

On 16 GB, such as an RTX 5080, Q3_K_M (13.4 GB) fits with an 8K context; IQ4_XS (15.5 GB) is just over the limit once the cache is added.^[4][6] On 12 GB, such as an RTX 3060, Q2_K (10.8 GB) is tight and IQ2_XS (9.1 GB) fits, with a clear quality loss at 2 bits.^[4][5]

How much VRAM does the full 256K context need?

At FP16, the KV cache for 262,144 tokens is about 17.2 GB.^[3] With Q4_K_M weights (17.4 GB), that's about 34.6 GB before engine overhead, which is more than a 32 GB RTX 5090. A 48 GB card handles it.

Can a Mac run Qwen3.8 27B?

Yes. A Mac with 32 GB of unified memory can hold Q6_K (23.9 GB) with a moderate context, and a 64 GB Mac can run Q8_0 (29.1 GB) or even the BF16 weights at a short context.^[4][11]

How we calculated

• Weights are the exact GGUF file sizes on Hugging Face, with multi-part files added together.^[4] BF16 is the two-part BF16 GGUF.
• KV cache is calculated from config.json for the 16 full-attention layers only. We assume an FP16 cache.^[3]
• Device memory is from NVIDIA and Apple spec pages. We compare in decimal GB, so the results lean conservative.
• Not included: engine working buffers, the fixed linear-attention state, the OS and display, and the vision projector. Leave headroom.

Sources

All sources checked on 2026-10-23.

1. Hugging Face, Qwen/Qwen3.8-27B (parameter count, license, repo date, downloads). https://huggingface.co/Qwen/Qwen3.8-27B
2. Hugging Face, Qwen/Qwen3.8-27B file list. https://huggingface.co/Qwen/Qwen3.8-27B/tree/main
3. Hugging Face, config.json (layers, attention layout, KV heads, context). https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/config.json
4. Hugging Face, bartowski/Qwen3.8-27B-GGUF file list. https://huggingface.co/bartowski/Qwen3.8-27B-GGUF/tree/main
5. NVIDIA, GeForce RTX 3060 family (12 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3060-3060ti/
6. NVIDIA, GeForce RTX 5080 (16 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5080/
7. NVIDIA, GeForce RTX 3090 specs (24 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090-3090ti/
8. NVIDIA, GeForce RTX 4090 specs (24 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/40-series/rtx-4090/
9. NVIDIA, GeForce RTX 5090 specs (32 GB). https://www.nvidia.com/en-us/geforce/graphics-cards/50-series/rtx-5090/
10. NVIDIA, RTX 6000 Ada Generation (48 GB). https://www.nvidia.com/en-us/products/workstations/rtx-6000/
11. Apple, MacBook Pro tech specs (unified memory options). https://www.apple.com/macbook-pro/specs/
12. Runpod, GPU pricing (checked October 23, 2026). https://www.runpod.io/pricing
13. Ollama, qwen3.8 library page. https://ollama.com/library/qwen3.8