# Can the Instinct MI50 32GB run GLiDE-27B?

> Yes. The Instinct MI50 32GB runs GLiDE-27B fully in VRAM at Q8_0 (about 30 GB of 32 GB), at roughly 20 tokens/s.

Instinct MI50 32GB: 32 GB (32 GB usable), 1024 GB/s. Estimates at an 8K context.

## Every quantisation

| Quant | Quality | ≈ GB | Fit | Est. tok/s |
|---|---|---|---|---|
| Q2_K | Heavy loss | 10.89 | fits | 66 |
| Q3_K_M | Noticeable loss | 14.97 | fits | 44 |
| Q4_K_M | Small loss | 18.68 | fits | 36 |
| Q5_K_M | Near-lossless | 20.64 | fits | 30 |
| Q6_K | Near-lossless | 23.51 | fits | 26 |
| Q8_0 (pick) | Lossless | 29.6 | fits | 20 |
| F16 | Lossless | 53.52 | offload | 1.4 |

## Run it

OpenAI-compatible server with a web UI (llama.cpp):

```bash
llama-server -hf seamon67/GLiDE-27B-GGUF:Q4_K_M
```

Docker with Vulkan (llama.cpp):

```bash
docker run -d --name llama --device /dev/dri -p 8080:8080 \
  -v ~/.cache/llama.cpp:/root/.cache/llama.cpp \
  ghcr.io/ggml-org/llama.cpp:server-vulkan \
  -hf seamon67/GLiDE-27B-GGUF:Q4_K_M --host 0.0.0.0 --port 8080 -ngl 99
```

Ask about an image (transformers):

```python
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="fastino/GLiDE-27B", dtype="auto", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
            {"type": "text", "text": "Describe this image."},
        ],
    }
]
out = pipe(text=messages, max_new_tokens=256)
print(out[0]["generated_text"][-1]["content"])
```


## Vision models that fit comfortably

- [google/gemma-4-26B-A4B-it](https://selfhost.directory/models/google/gemma-4-26b-a4b-it/on/instinct-mi50): Q8_0 ≈28.65 GB, est. 133 tok/s
- [Qwen/Qwen3.8-27B](https://selfhost.directory/models/qwen/qwen3.8-27b/on/instinct-mi50): Q6_K ≈24.88 GB, est. 25 tok/s
- [Qwen/Qwen3.6-27B](https://selfhost.directory/models/qwen/qwen3.6-27b/on/instinct-mi50): Q6_K ≈24.53 GB, est. 25 tok/s
- [google/gemma-4-31B-it](https://selfhost.directory/models/google/gemma-4-31b-it/on/instinct-mi50): Q5_K_M ≈29.18 GB, est. 25 tok/s
- [google/gemma-3-27b-it](https://selfhost.directory/models/google/gemma-3-27b-it/on/instinct-mi50): Q6_K ≈26.05 GB, est. 25 tok/s
- [Qwen/Qwen3.5-27B](https://selfhost.directory/models/qwen/qwen3.5-27b/on/instinct-mi50): Q6_K ≈24.46 GB, est. 25 tok/s

## Other GPUs that run it

- [Tesla T4](https://selfhost.directory/models/fastino/glide-27b/on/tesla-t4): Q3_K_M, est. 14 tok/s
- [Arc A770 16GB](https://selfhost.directory/models/fastino/glide-27b/on/arc-a770): Q3_K_M, est. 24 tok/s
- [Radeon RX 7800 XT](https://selfhost.directory/models/fastino/glide-27b/on/rx-7800-xt): Q3_K_M, est. 27 tok/s
- [Tesla P40](https://selfhost.directory/models/fastino/glide-27b/on/tesla-p40): Q5_K_M, est. 10 tok/s
- [Radeon RX 7900 XTX](https://selfhost.directory/models/fastino/glide-27b/on/rx-7900-xtx): Q5_K_M, est. 28 tok/s
- [GeForce RTX 5090](https://selfhost.directory/models/fastino/glide-27b/on/rtx-5090): Q8_0, est. 36 tok/s
- [Apple M4 Pro (48 GB)](https://selfhost.directory/models/fastino/glide-27b/on/apple-m4-pro-48gb): Q8_0, est. 5.4 tok/s
- [RTX A6000](https://selfhost.directory/models/fastino/glide-27b/on/rtx-a6000): Q8_0, est. 15 tok/s

---
Model: https://selfhost.directory/models/fastino/glide-27b. GPU: https://selfhost.directory/models/gpus/instinct-mi50. Source: https://selfhost.directory/models/fastino/glide-27b/on/instinct-mi50