Qwen3.8-27B GGUF: which quant fits your GPU, and how much context?

Qwen3.8-27B at standard 4-bit (the 16.46 GB UD-Q4_K_M file) needs about 17.80 GB with 8,192 tokens of context, so it fits a 24 GB card with room for about 91K tokens. For the full 262,144-token window at 4-bit you need about 34.44 GB, more than one 24 GB or 32 GB card gives at Q4_K_M with an FP16 cache.

Every figure below is our estimate from the model's config and the published file sizes, not a measurement.

How much VRAM does Qwen3.8-27B need?

The total is the file, plus 65,536 bytes per token for an FP16 cache (16 full-attention layers × 4 KV heads × 256 × 2 × 2 bytes), plus 0.80 GB of working space. A card keeps 0.6 GB free, so 12 GB leaves 11.4 GB, 16 GB leaves 15.4 GB, 24 GB leaves 23.4 GB and 32 GB leaves 31.4 GB. Work a different length in the VRAM calculator.

How much context fits on 12, 16, 24 and 32 GB?

“no fit” means the file does not fit even at 8,192 tokens. “262K (max)” means the model's 262,144-token limit is reached before memory runs out. The image projector (mmproj, 0.93 GB) and the MTP drafter (1.37 GB) are not in these totals. On 24 GB, UD-Q4_K_M plus both of those leaves about 57K tokens instead of 91K.

Does a q8_0 cache double the context?

A q8_0 cache (`-ctk q8_0 -ctv q8_0` in llama.cpp) halves conversation memory. The table is our arithmetic. The effect on answer quality is not measured.

Why the cache is so small: 16 of 64 layers

Only 16 of Qwen3.8-27B's 64 layers use full attention, with 4 KV heads of dimension 256. That is about 65.5 KB per token. The other 48 layers use linear attention (Gated DeltaNet). Their fixed-size state does not grow with context, and that state size is not in our estimate.

Which card should I buy for long context?

The same model on one card: Can I run Qwen3.8-27B. What else fits each size: 12 GB, 16 GB, 24 GB and 32 GB. The used 24 GB card we price: GeForce RTX 3090 24GB.

Community measurement, not ours (published 18 Sept 2026): on one RTX 3090, llama.cpp as shipped about 27.6 tokens per second, and llama.cpp tuned (flash attention, a q8 KV cache and MTP) about 42.6 tokens per second, on a 4-bit checkpoint with a prompt of about 1,000 tokens. Speed at a long context is not measured. Source.

Compare

Qwen3.8-27B next to other models: Qwen3.8-27B vs Gemma 4 31B, Qwen3.6-35B-A3B vs Qwen3.8-27B, and what the quant names mean. The model page is Qwen3.8-27B, and the fit page is Can I run Qwen3.8-27B.

Sources

Config: Qwen/Qwen3.8-27B config.json. File sizes, checked 24 Oct 2026: unsloth/Qwen3.8-27B-GGUF. Our model row: /api/models. Stored prices: /api/hardware.