At 8,192 tokens, Qwen3-8B Q4_K_M stores 1.13 GiB of FP16 KV because every layer keeps the full context, and the total is 5.93 to 6.48 GiB. Qwen3.8-27B Q4_K_M stores 0.50 GiB of attention KV plus 0.14 GiB of Gated DeltaNet state. Counting all 64 layers with the standard GQA formula (4 KV heads, head dim 256) would be 2.00 GiB of KV. The calculator does not do that. Coverage is partial, so the published total includes a conservative pad on the high bound: 16.43 to 19.74 GiB. Formula estimates, calculator 0.3.0, dataset 2026-10-02.1, llama.cpp, 1 sequence, FP16 KV. No calibration and no planned benchmark for Qwen3.8-27B. Software support is unverified. Tokens per second are Unknown.
The selected file is Qwen3.8-27B.Q4_K_M.gguf from mradermacher/Qwen3.8-27B-GGUF, 16,810,714,848 bytes, weights 15.66 GiB. The architecture record is 16 groups, one gated-attention layer and three Gated DeltaNet layers per group, recurrent state dtype fp32. That is 16 attention layers and 48 recurrent layers.
Three layouts at 8,192 and 32,768 tokens
No device is selected. Qwen3-8B and Gemma 3 27B are full coverage. Qwen3.8-27B is partial.
| Checkpoint | File | KV at 8,192 | Other state at 8,192 | Total at 8,192 | KV at 32,768 | Total at 32,768 |
|---|---|---|---|---|---|---|
| Qwen3-8B Q4_K_M | 5,027,783,488 | 1.13 | 0 | 5.93 to 6.48 | 4.50 | 9.31 to 10.09 |
| Gemma 3 27B Q4_0 | 17,229,630,496 | 1.03 | 0 | 17.20 to 17.97 | 2.91 | 19.08 to 19.98 |
| Qwen3.8-27B Q4_K_M | 16,810,714,848 | 0.50 | 0.14 | 16.43 to 19.74 | 2.00 | 17.93 to 21.58 |
Qwen3-8B KV quadruples from 8,192 to 32,768 tokens. Qwen3.8 attention KV also quadruples, from 0.50 GiB to 2.00 GiB, and the 0.14 GiB recurrent state does not. Gemma 3 27B moves from 1.03 GiB to 2.91 GiB because of its sliding window, which is full coverage and has no conservative pad. The pad is why Qwen3.8’s high bound jumps further than its KV term. The unpadded whole-model range at 8,192 tokens is 16.43 to 17.16 GiB. The number this calculator shows on the result is the padded total.
Desktop apps sometimes preview memory with a file size or with KV on every block. This page does not quote those previews. It shows the three catalog totals above. Gemma’s window is written up on Gemma 3 sliding-window VRAM. The dense Qwen3 sizes are on Qwen3 VRAM by size. The Qwen3.8 model page is Qwen3.8-27B. The cache tool is the KV cache calculator.
FAQ
Why is Qwen3.8-27B KV smaller than a 64-layer GQA count?
The config has 64 layers, 4 KV heads and head dim 256. Applying the standard formula to all 64 layers at 8,192 tokens is 2.00 GiB. The adapter uses 16 gated-attention layers, so attention KV is 0.50 GiB, plus 0.14 GiB of Gated DeltaNet state.
What is the Qwen3.8-27B Q4 total?
mradermacher Qwen3.8-27B.Q4_K_M.gguf is 16,810,714,848 bytes. At 8,192 tokens the calculator total is 16.43 to 19.74 GiB. Coverage is partial, so the high end includes a conservative pad. There is no calibration and no planned benchmark for this checkpoint.
Does the recurrent state grow with context?
Not in this estimate. At 32,768 tokens attention KV rises to 2.00 GiB and the Gated DeltaNet state stays 0.14 GiB. The total is 17.93 to 21.58 GiB, still partial coverage.