Qwen3.8-27B GGUF size and VRAM requirements
unsloth/Qwen3.8-27B-GGUF is a large language model with 27.32 billion parameters, built on the qwen35 architecture. It is released under the apache-2.0 license and has been downloaded 867,963 times.
Under the hood it uses 64 transformer layers, a hidden size of 5,120, 24 attention heads. It uses a hybrid attention design: only 16 of its 64 layers keep a full KV cache, while the rest use linear attention with a constant-size state. Long contexts therefore cost far less memory than on a conventional model of this size.
To run unsloth/Qwen3.8-27B-GGUF locally at a 4,096-token context, its quantized versions need between 9.44 GB (IQ2_XXS, lowest quality) and 51.95 GB (BF16, highest quality) of memory, weights plus KV cache and a system margin included. Context length drives that memory directly: at 4,096 tokens the KV cache for Qwen3.8-27B-GGUF is about 0.25 GB, rising to roughly 16.0 GB at its full 262,144-token context. Shorter prompts free up memory for a higher-quality quantization.
For most users the best balance is Q6_K, needing about 22.36 GB. That means unsloth/Qwen3.8-27B-GGUF fits entirely in the VRAM of a 10 GB GPU or larger, running fully on the GPU.
Available GGUF quantizations for unsloth/Qwen3.8-27B-GGUF include IQ2_XXS, IQ2_M, Q2_K_XL, IQ3_XXS, Q3_K_S, Q3_K_XL, Q3_K_M, IQ4_XS, Q4_0, Q4_K_S, IQ4_NL, Q4_K_M, Q4_1, Q4_K_XL, Q5_K_S, Q5_K_M, Q5_K_XL, Q6_K, Q6_K_XL, Q8_0, Q8_K_XL, BF16. The model supports a native context length of up to 262,144 tokens; a longer context grows the KV cache and the memory needed.
GGUF file size and memory by quantization
Compare real GGUF weight sizes, estimated KV cache and total memory for Q4, Q5, Q8 and every quantization published in this repository.
| Quant. | Bits | Quality | Weights | KV | Total | Speed~ | Verdict |
|---|---|---|---|---|---|---|---|
| IQ2_XXS | 2.64 | Low | 8.39 GB | 0.25 GB | 9.44 GB | 6.0 t/s | Offload |
| IQ2_M | 3.02 | Low | 9.61 GB | 0.25 GB | 10.66 GB | 5.2 t/s | Offload |
| Q2_K_XL | 3.13 | Low | 9.94 GB | 0.25 GB | 10.99 GB | 5.0 t/s | Offload |
| IQ3_XXS | 3.49 | Fair | 11.1 GB | 0.25 GB | 12.15 GB | 4.5 t/s | Offload |
| Q3_K_S | 3.68 | Fair | 11.71 GB | 0.25 GB | 12.76 GB | 4.3 t/s | Offload |
| Q3_K_XL | 3.94 | Fair | 12.52 GB | 0.25 GB | 13.57 GB | 4.0 t/s | Offload |
| Q3_K_M | 4.05 | Fair | 12.87 GB | 0.25 GB | 13.92 GB | 3.9 t/s | Offload |
| IQ4_XS | 4.6 | Good | 14.63 GB | 0.25 GB | 15.68 GB | 3.4 t/s | Offload |
| Q4_0 | 4.7 | Good | 14.95 GB | 0.25 GB | 16.0 GB | 3.3 t/s | Offload |
| Q4_K_S | 4.72 | Good | 15.01 GB | 0.25 GB | 16.06 GB | 3.3 t/s | Offload |
| IQ4_NL | 4.78 | Good | 15.22 GB | 0.25 GB | 16.27 GB | 3.3 t/s | Offload |
| Q4_K_M | 5.01 | Very good | 15.93 GB | 0.25 GB | 16.98 GB | 3.1 t/s | Offload |
| Q4_1 | 5.14 | Very good | 16.34 GB | 0.25 GB | 17.39 GB | 3.1 t/s | Offload |
| Q4_K_XL | 5.25 | Very good | 16.69 GB | 0.25 GB | 17.74 GB | 3.0 t/s | Offload |
| Q5_K_S | 5.64 | Very good | 17.95 GB | 0.25 GB | 19.0 GB | 2.8 t/s | Offload |
| Q5_K_M | 5.81 | Very good | 18.47 GB | 0.25 GB | 19.52 GB | 2.7 t/s | Offload |
| Q5_K_XL | 5.92 | Very good | 18.83 GB | 0.25 GB | 19.88 GB | 2.7 t/s | Offload |
| Q6_K | 6.7 | Excellent | 21.31 GB | 0.25 GB | 22.36 GB | 2.3 t/s | Offload |
| Q6_K_XL | 7.59 | Excellent | 24.14 GB | 0.25 GB | 25.19 GB | — | Insufficient |
| Q8_0 | 8.51 | Excellent | 27.05 GB | 0.25 GB | 28.1 GB | — | Insufficient |
| Q8_K_XL | 9.21 | Excellent | 29.3 GB | 0.25 GB | 30.35 GB | — | Insufficient |
| BF16 | 16.0 | Excellent | 50.9 GB | 0.25 GB | 51.95 GB | — | Insufficient |
KV cache computed from the model's exact architecture. Speed is a rough estimate bounded by memory bandwidth.
Frequently asked questions
What kind of model is unsloth/Qwen3.8-27B-GGUF?
unsloth/Qwen3.8-27B-GGUF is a language model with 27.32 billion parameters, based on the qwen35 architecture. It is released under the apache-2.0 license and distributed as GGUF files for local inference.
How does unsloth/Qwen3.8-27B-GGUF handle long context?
unsloth/Qwen3.8-27B-GGUF uses a hybrid attention design: only 16 of its 64 layers keep a full KV cache, and the rest use linear attention with a constant-size state. Long contexts cost far less memory than on a conventional model of this size.
How much VRAM do you need to run unsloth/Qwen3.8-27B-GGUF?
You need about 9.44 GB of VRAM to run unsloth/Qwen3.8-27B-GGUF entirely on the GPU using the IQ2_XXS quantization (at a 4,096-token context). Smaller quantizations lower the requirement at the cost of quality.
Can I run unsloth/Qwen3.8-27B-GGUF on an 8 GB GPU?
Partially. unsloth/Qwen3.8-27B-GGUF only fits on an 8 GB GPU by offloading part of it to system RAM (with Q6_K), which runs but is slower.
Can I run unsloth/Qwen3.8-27B-GGUF on a 16 GB GPU?
Yes. With 16 GB of VRAM you can run unsloth/Qwen3.8-27B-GGUF fully on the GPU using IQ4_XS (about 15.68 GB).
Can I run unsloth/Qwen3.8-27B-GGUF on a 24 GB GPU?
Yes. With 24 GB of VRAM you can run unsloth/Qwen3.8-27B-GGUF fully on the GPU using Q6_K (about 22.36 GB).
What context length does unsloth/Qwen3.8-27B-GGUF support?
unsloth/Qwen3.8-27B-GGUF supports a native context length of up to 262,144 tokens. A longer context grows the KV cache, so it increases the memory needed to run the model.
What is the best quantization for unsloth/Qwen3.8-27B-GGUF?
For unsloth/Qwen3.8-27B-GGUF, a strong default is Q4_K_M, which needs about 16.98 GB and keeps most of the quality while roughly halving the memory versus 8-bit. With VRAM to spare, Q5_K_M or Q6_K add a little more quality; if you are tight on memory, a smaller quantization still runs. Pick the highest quantization that fits your VRAM.