STARK-WEB-12B-v1.7-i1 GGUF size and VRAM requirements
mradermacher/STARK-WEB-12B-v1.7-i1-GGUF is a large language model with 11.91 billion parameters, built on the gemma4 architecture. It is released under the apache-2.0 license and has been downloaded 11,633 times.
To run mradermacher/STARK-WEB-12B-v1.7-i1-GGUF locally at a 4,096-token context, its quantized versions need between 2.47 GB (GGUF, lowest quality) and 11.58 GB (Q6_K, highest quality) of memory, weights plus KV cache and a system margin included.
For most users the best balance is IQ3_M, needing about 7.81 GB. That means mradermacher/STARK-WEB-12B-v1.7-i1-GGUF fits entirely in the VRAM of a 6 GB GPU or larger, running fully on the GPU.
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 |
|---|---|---|---|---|---|---|---|
| GGUF | 0.01 | Very low | 0.01 GB | 1.67 GB | 2.47 GB | 57414.7 t/s | Fits in VRAM |
| IQ1_S | 2.0 | Very low | 2.78 GB | 1.67 GB | 5.24 GB | 143.9 t/s | Fits in VRAM |
| IQ1_M | 2.15 | Very low | 2.98 GB | 1.67 GB | 5.45 GB | 134.1 t/s | Fits in VRAM |
| IQ2_XXS | 2.4 | Very low | 3.32 GB | 1.67 GB | 5.79 GB | 120.4 t/s | Fits in VRAM |
| IQ2_XS | 2.61 | Low | 3.62 GB | 1.67 GB | 6.09 GB | 110.5 t/s | Fits in VRAM |
| IQ2_S | 2.74 | Low | 3.8 GB | 1.67 GB | 6.26 GB | 105.3 t/s | Fits in VRAM |
| IQ2_M | 2.94 | Low | 4.07 GB | 1.67 GB | 6.54 GB | 98.2 t/s | Fits in VRAM |
| Q2_K_S | 3.03 | Low | 4.19 GB | 1.67 GB | 6.66 GB | 95.4 t/s | Fits in VRAM |
| Q2_K | 3.25 | Low | 4.5 GB | 1.67 GB | 6.96 GB | 88.9 t/s | Fits in VRAM |
| IQ3_XXS | 3.26 | Low | 4.52 GB | 1.67 GB | 6.98 GB | 88.6 t/s | Fits in VRAM |
| IQ3_XS | 3.54 | Fair | 4.91 GB | 1.67 GB | 7.38 GB | 81.5 t/s | Fits in VRAM |
| IQ3_S | 3.71 | Fair | 5.15 GB | 1.67 GB | 7.61 GB | 77.7 t/s | Fits in VRAM |
| Q3_K_S | 3.71 | Fair | 5.15 GB | 1.67 GB | 7.61 GB | 77.7 t/s | Fits in VRAM |
| IQ3_M | 3.85 | Fair | 5.34 GB | 1.67 GB | 7.81 GB | 74.9 t/s | Fits in VRAM |
| Q3_K_M | 4.09 | Fair | 5.67 GB | 1.67 GB | 8.13 GB | 8.8 t/s | Offload |
| Q3_K_L | 4.41 | Good | 6.12 GB | 1.67 GB | 8.58 GB | 8.2 t/s | Offload |
| IQ4_XS | 4.46 | Good | 6.18 GB | 1.67 GB | 8.64 GB | 8.1 t/s | Offload |
| IQ4_NL | 4.69 | Good | 6.5 GB | 1.67 GB | 8.96 GB | 7.7 t/s | Offload |
| Q4_0 | 4.7 | Good | 6.52 GB | 1.67 GB | 8.98 GB | 7.7 t/s | Offload |
| Q4_K_S | 4.72 | Good | 6.54 GB | 1.67 GB | 9.01 GB | 7.6 t/s | Offload |
| Q4_K_M | 4.96 | Good | 6.87 GB | 1.67 GB | 9.34 GB | 7.3 t/s | Offload |
| Q4_1 | 5.14 | Very good | 7.13 GB | 1.67 GB | 9.6 GB | 7.0 t/s | Offload |
| Q5_K_S | 5.6 | Very good | 7.77 GB | 1.67 GB | 10.23 GB | 6.4 t/s | Offload |
| Q5_K_M | 5.74 | Very good | 7.96 GB | 1.67 GB | 10.43 GB | 6.3 t/s | Offload |
| Q6_K | 6.57 | Excellent | 9.11 GB | 1.67 GB | 11.58 GB | 5.5 t/s | Offload |
KV cache estimated (architecture unavailable). Speed is a rough estimate bounded by memory bandwidth.
Frequently asked questions
How much VRAM do you need to run mradermacher/STARK-WEB-12B-v1.7-i1-GGUF?
You need about 5.79 GB of VRAM to run mradermacher/STARK-WEB-12B-v1.7-i1-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 mradermacher/STARK-WEB-12B-v1.7-i1-GGUF on an 8 GB GPU?
Yes. With 8 GB of VRAM you can run mradermacher/STARK-WEB-12B-v1.7-i1-GGUF fully on the GPU using IQ3_M (about 7.81 GB).
Can I run mradermacher/STARK-WEB-12B-v1.7-i1-GGUF on a 16 GB GPU?
Yes. With 16 GB of VRAM you can run mradermacher/STARK-WEB-12B-v1.7-i1-GGUF fully on the GPU using Q6_K (about 11.58 GB).
Can I run mradermacher/STARK-WEB-12B-v1.7-i1-GGUF on a 24 GB GPU?
Yes. With 24 GB of VRAM you can run mradermacher/STARK-WEB-12B-v1.7-i1-GGUF fully on the GPU using Q6_K (about 11.58 GB).
What is the best quantization for mradermacher/STARK-WEB-12B-v1.7-i1-GGUF?
If memory allows, higher bits-per-weight means better quality. A common sweet spot is a Q4_K_M or Q5_K_M quantization, which keeps most of the quality while roughly halving the memory versus 8-bit. Pick the highest quantization that still fits in your VRAM.