gpt-oss-20b GGUF size and VRAM requirements
unsloth/gpt-oss-20b-GGUF is a large language model with 20.91 billion parameters, built on the gpt-oss architecture. It is released under the apache-2.0 license and has been downloaded 541,824 times.
gpt-oss-20b-GGUF is a Mixture-of-Experts model with 32 experts, of which 4 are active on each token. Routing only 4 of 32 experts makes it noticeably faster than a dense model of the same size — but every expert still has to be held in memory, so the numbers below are set by the full parameter count, not the active one. It is built from 24 transformer layers, a hidden size of 2,880, 64 attention heads. It uses a hybrid attention design: only 12 of its 24 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/gpt-oss-20b-GGUF locally at a 4,096-token context, its quantized versions need between 11.57 GB (Q3_K_S, lowest quality) and 13.74 GB (F16, 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 gpt-oss-20b-GGUF is about 0.09 GB, rising to roughly 3.0 GB at its full 131,072-token context. Shorter prompts free up memory for a higher-quality quantization.
For most users the best balance is F16, needing about 13.74 GB. That means unsloth/gpt-oss-20b-GGUF fits entirely in the VRAM of a 12 GB GPU or larger, running fully on the GPU.
Available GGUF quantizations for unsloth/gpt-oss-20b-GGUF include Q3_K_S, Q2_K, Q4_0, Q3_K_M, Q4_1, Q4_K_S, Q4_K_M, Q5_K_S, Q5_K_M, Q2_K_L, Q4_K_XL, Q6_K, Q6_K_XL, Q8_0, Q8_K_XL, F16. The model supports a native context length of up to 131,072 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 |
|---|---|---|---|---|---|---|---|
| Q3_K_S | 4.38 | Good | 10.68 GB | 0.09 GB | 11.57 GB | 4.7 t/s | Offload |
| Q2_K | 4.39 | Good | 10.68 GB | 0.09 GB | 11.57 GB | 4.7 t/s | Offload |
| Q4_0 | 4.4 | Good | 10.71 GB | 0.09 GB | 11.61 GB | 4.7 t/s | Offload |
| Q3_K_M | 4.4 | Good | 10.72 GB | 0.09 GB | 11.61 GB | 4.7 t/s | Offload |
| Q4_1 | 4.43 | Good | 10.78 GB | 0.09 GB | 11.68 GB | 4.6 t/s | Offload |
| Q4_K_S | 4.44 | Good | 10.82 GB | 0.09 GB | 11.71 GB | 4.6 t/s | Offload |
| Q4_K_M | 4.45 | Good | 10.83 GB | 0.09 GB | 11.72 GB | 4.6 t/s | Offload |
| Q5_K_S | 4.48 | Good | 10.91 GB | 0.09 GB | 11.8 GB | 4.6 t/s | Offload |
| Q5_K_M | 4.48 | Good | 10.91 GB | 0.09 GB | 11.81 GB | 4.6 t/s | Offload |
| Q2_K_L | 4.5 | Good | 10.95 GB | 0.09 GB | 11.84 GB | 4.6 t/s | Offload |
| Q4_K_XL | 4.54 | Good | 11.06 GB | 0.09 GB | 11.95 GB | 4.5 t/s | Offload |
| Q6_K | 4.61 | Good | 11.21 GB | 0.09 GB | 12.11 GB | 4.5 t/s | Offload |
| Q6_K_XL | 4.61 | Good | 11.21 GB | 0.09 GB | 12.11 GB | 4.5 t/s | Offload |
| Q8_0 | 4.63 | Good | 11.28 GB | 0.09 GB | 12.17 GB | 4.4 t/s | Offload |
| Q8_K_XL | 5.05 | Very good | 12.29 GB | 0.09 GB | 13.18 GB | 4.1 t/s | Offload |
| F16 | 5.28 | Very good | 12.85 GB | 0.09 GB | 13.74 GB | 3.9 t/s | Offload |
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/gpt-oss-20b-GGUF?
unsloth/gpt-oss-20b-GGUF is a language model with 20.91 billion parameters, based on the gpt-oss architecture. It is released under the apache-2.0 license and distributed as GGUF files for local inference.
Is unsloth/gpt-oss-20b-GGUF a Mixture-of-Experts (MoE) model?
Yes. unsloth/gpt-oss-20b-GGUF is a Mixture-of-Experts model with 32 experts, of which 4 are activated per token. That makes it faster than a dense model of the same size, but all 32 experts must be loaded into memory, so the VRAM/RAM it needs is driven by the total parameter count, not the active one.
How does unsloth/gpt-oss-20b-GGUF handle long context?
unsloth/gpt-oss-20b-GGUF uses a hybrid attention design: only 12 of its 24 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/gpt-oss-20b-GGUF?
You need about 11.95 GB of VRAM to run unsloth/gpt-oss-20b-GGUF entirely on the GPU using the Q4_K_XL quantization (at a 4,096-token context). Smaller quantizations lower the requirement at the cost of quality.
Can I run unsloth/gpt-oss-20b-GGUF on an 8 GB GPU?
Partially. unsloth/gpt-oss-20b-GGUF only fits on an 8 GB GPU by offloading part of it to system RAM (with F16), which runs but is slower.
Can I run unsloth/gpt-oss-20b-GGUF on a 16 GB GPU?
Yes. With 16 GB of VRAM you can run unsloth/gpt-oss-20b-GGUF fully on the GPU using F16 (about 13.74 GB).
Can I run unsloth/gpt-oss-20b-GGUF on a 24 GB GPU?
Yes. With 24 GB of VRAM you can run unsloth/gpt-oss-20b-GGUF fully on the GPU using F16 (about 13.74 GB).
What context length does unsloth/gpt-oss-20b-GGUF support?
unsloth/gpt-oss-20b-GGUF supports a native context length of up to 131,072 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/gpt-oss-20b-GGUF?
For unsloth/gpt-oss-20b-GGUF, a strong default is Q4_K_M, which needs about 11.72 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.