This page guesses your card from what the browser exposes; the desktop app measures it with the driver, and the app is the thing that also runs the model. Treat every number below as a good second opinion, not a reading.
Your machine
reading…- Graphics card
- —
- VRAM
- —
- RAM
- —
- CPU cores
- —
- WebGPU
- —
- Download room
- —
Set your card
Browsers do not expose VRAM, so the figure above comes from recognising the card's name. If that name is wrong, or your browser would not give one, pick the card here. A result built from your pick is labelled manual and is never shown as detected.
What fits
Working out what your machine will hold…
The 10 models this page checks
Download sizes are the GGUF file sizes in gigabytes, from the desktop app's model catalog (version 1, updated 2026-10-20). A model needs its file size plus room for its context (the KV cache) and about 0.5 GB of overhead, so a card needs a little more memory than the file. The checker above does that sum for your card.
| Model | Good for | Download size by quant | License |
|---|---|---|---|
| Qwen3.8-27B | reasoning | UD-Q4_K_M 16.46 GB, UD-Q5_K_M 19.77 GB, UD-Q6_K 21.98 GB, Q8_0 29.05 GB | Apache-2.0 |
| Qwen3-32B | reasoning | Q4_K_M 19.76 GB, Q5_K_M 23.21 GB, Q6_K 26.88 GB, Q8_0 34.82 GB | Apache-2.0 |
| Qwen3-Coder-30B-A3B-Instruct | coding | Q4_K_M 18.56 GB, Q5_K_M 21.73 GB, Q6_K 25.09 GB, Q8_0 32.48 GB | Apache-2.0 |
| Qwen3-30B-A3B | chat | Q4_K_M 18.56 GB, Q5_K_M 21.73 GB, Q8_0 32.48 GB | Apache-2.0 |
| Qwen3-14B | chat | Q4_K_M 9 GB, Q6_K 12.12 GB, Q8_0 15.7 GB | Apache-2.0 |
| Qwen3-8B | chat | Q4_K_M 5.03 GB, Q5_K_M 5.85 GB, Q8_0 8.71 GB | Apache-2.0 |
| Qwen3-4B | chat | Q4_K_M 2.5 GB, Q5_K_M 2.89 GB, Q8_0 4.28 GB | Apache-2.0 |
| Qwen3-VL-8B-Instruct | images + text | Q4_K_M 5.03 GB, Q8_0 8.71 GB | Apache-2.0 |
| Qwen3-Embedding-0.6B | search (embeddings) | Q8_0 0.64 GB, F16 1.2 GB | Apache-2.0 |
| Llama-3.3-70B-Instruct | chat | Q4_K_M 42.52 GB, Q8_0 74.98 GB | Llama 3.3 Community License |
What fits on an 8, 12, 16, 24 or 32 GB card
The same arithmetic the checker runs, worked out ahead of time for common amounts of graphics memory (as of October 2026). Each model is shown at the biggest quant that fits fully on the card, with its weights, its KV cache at the context it is served with and about 0.5 GB of runtime overhead.
| Card memory | Models that fit | Which ones, at what quant | The rest |
|---|---|---|---|
| 8 GB | 2 of 10 | Qwen3-4B (Q8_0, needs 6.7 GB); Qwen3-Embedding-0.6B (F16, needs 2.5 GB) | 8 need more memory or CPU offload |
| 12 GB | 4 of 10 | Qwen3-8B (Q5_K_M, needs 10.4 GB); Qwen3-VL-8B-Instruct (Q4_K_M, needs 10.4 GB); Qwen3-4B (Q8_0, needs 6.7 GB); Qwen3-Embedding-0.6B (F16, needs 2.5 GB) | 6 need more memory or CPU offload |
| 16 GB | 5 of 10 | Qwen3-14B (Q4_K_M, needs 13.9 GB); Qwen3-VL-8B-Instruct (Q8_0, needs 13.8 GB); Qwen3-8B (Q8_0, needs 13.1 GB); Qwen3-4B (Q8_0, needs 6.7 GB); Qwen3-Embedding-0.6B (F16, needs 2.5 GB) | 5 need more memory or CPU offload |
| 24 GB | 9 of 10 | Qwen3-Coder-30B-A3B-Instruct (Q5_K_M, needs 23.7 GB); Qwen3-30B-A3B (Q5_K_M, needs 23.7 GB); Qwen3.8-27B (UD-Q5_K_M, needs 22.9 GB); Qwen3-32B (Q4_K_M, needs 20.9 GB); Qwen3-14B (Q8_0, needs 20.1 GB); Qwen3-VL-8B-Instruct (Q8_0, needs 13.8 GB); Qwen3-8B (Q8_0, needs 13.1 GB); Qwen3-4B (Q8_0, needs 6.7 GB); Qwen3-Embedding-0.6B (F16, needs 2.5 GB) | 1 needs more memory or CPU offload |
| 32 GB | 9 of 10 | Qwen3.8-27B (Q8_0, needs 31.6 GB); Qwen3-32B (Q6_K, needs 27.5 GB); Qwen3-Coder-30B-A3B-Instruct (Q6_K, needs 26.9 GB); Qwen3-30B-A3B (Q5_K_M, needs 23.7 GB); Qwen3-14B (Q8_0, needs 20.1 GB); Qwen3-VL-8B-Instruct (Q8_0, needs 13.8 GB); Qwen3-8B (Q8_0, needs 13.1 GB); Qwen3-4B (Q8_0, needs 6.7 GB); Qwen3-Embedding-0.6B (F16, needs 2.5 GB) | 1 needs more memory or CPU offload |
No tokens per second, on purpose
Speed depends on the model, the quant, the driver, the context length and whatever else is using the GPU, and a page that has not loaded the model cannot measure any of it. A made-up number would be worse than none, so this page does not show one. The desktop app runs the model, and can tell you what it actually did.
How these numbers are made
- Download size is the GGUF file size from the app's own model catalog, in decimal gigabytes as Hugging Face reports it.
- KV cache is 2 × layers × context × kvDim × 2 bytes, at the context the app would serve the model with, not the length the model was trained for.
- 0.5 GiB of runtime overhead covers the compute buffer and context initialisation.
- A model fits when that total is at or under your card's VRAM. It runs tight when it is over VRAM but under your RAM, which means CPU offload.
- The arithmetic is the desktop app's own, ported without changes, and a test pins the page to the app's byte results.