Which GPU do you need for local LLMs and image/video models?
Check whether a model fits a card at a given quantisation and context length, then see what renting that card costs. Every number carries an evidence grade.
Quick fit check
Result: Fits
- Weights: 17.8 GB · KV cache: 0.27 GB · Overhead: 2 GB
- Total: 20.1 GB · RTX 4090 24GB 24 GB · Headroom: 3.9 GB
VRAM by model
Minimum and recommended VRAM at the quantisation in each row. Section numbers (§) refer to the Chinese original, the local LLM hardware guide.
Text models
| Model | Params | Quant | Weights (GB) | Min VRAM (GB) | Recommended (GB) | Evidence | Source § |
|---|---|---|---|---|---|---|---|
| Qwen3-8B | 8B | Q4 | 5 | 5 | — | A | §2.1 |
| Qwen3-14B | 14B | Q4 | 9 | 9 | — | A | §2.1 |
| Qwen3-32B | 32B | Q4_K_M | 20 | 20 | — | A | §2.1 |
| Qwen3-30B-A3B | 30B | Q4 | not measured | not measured | — | A | §2.1 |
| Qwen3-235B-A22B | 235B | Q4 | not measured | single-card: no | — | A | §2.1 |
| Qwen3.8-27B | 27B | Q4_K_M | 17.8 | 17 | 24 | B | §2.2 |
Show all 4 rows
| Model | Params | Quant | Weights (GB) | Min VRAM (GB) | Recommended (GB) | Evidence | Source § |
|---|---|---|---|---|---|---|---|
| Qwen3.8-27B | 27B | Q3_K_M | 14 | 16 | — | B | §2.2 |
| Qwen3.8-27B | 27B | Q3_K_XL | 13 | 16 | — | B | §2.2 |
| Qwen3.8-27B | 27B | UD-Q2_K_XL | 10.7 | not measured | — | B | §2.2 |
| Qwen3.8-Max | 2400B | API | not measured | single-card: no | — | A | §2.2 |
Image models
| Model | Params | Quant | Weights (GB) | Min VRAM (GB) | Recommended (GB) | Evidence | Source § |
|---|---|---|---|---|---|---|---|
| Flux.2 Klein 4B | 4B | FP16 | 13 | 13 | — | A | §3.1 |
| Flux.2 Klein 9B | 9B | FP16 | 29 | 29 | — | A | §3.1 |
| Flux.2 Klein 9B | 9B | FP8 | 15 | 15 | — | A | §3.1 |
| Flux.2 Dev | — | FP16 | not measured | not measured | — | A | §3.1 |
| Flux.2 Dev | — | GGUF-Q4 | 19 | 19 | — | A | §3.1 |
Video models
| Model | Params | Quant | Weights (GB) | Min VRAM (GB) | Recommended (GB) | Evidence | Source § |
|---|---|---|---|---|---|---|---|
| Wan2.2-TI2V-5B | 5B | FP16 | not measured | 27.4 | — | B | §3.2 |
| Wan2.1/2.2-14B | 14B | FP16 | not measured | 65 | — | C | §3.2 |
| Wan2.1/2.2-14B | 14B | GGUF-Q5+T5-offload | not measured | not measured | — | C | §3.2 |
| Wan2.7-14B | 14B | Q5_K_M | not measured | 16 | 24 | C | §3.2 |
| Hailuo 3.0 (MiniMax H3) | — | INT8 | not measured | 21 | — | C | §3.3 |
GPU rental prices
Only the RTX 4090 has a published price. Other cards show 'not priced yet' instead of an extrapolated number.
Vast.ai market reference: on-demand single-GPU median over a rolling 7-day window (a day counts only with at least 20 offers). Cards without enough offers show the sample count instead of a number. Updated 2026-10-21.
| GPU | VRAM | Our price (instant / queued) | Host payout | Power cost, CN | Vast.ai market median (7-day) |
|---|---|---|---|---|---|
| RTX 4090 24GB | 24 GB | $0.25 / $0.15 | $0.20 | $0.05 | $0.54 |
| RTX 5090 32GB | 32 GB | Not priced yet | $0.22 | $0.06 | $0.75 |
| RTX 3090 24GB | 24 GB | Not priced yet | $0.08 | $0.04 | $0.22 |
| RTX 4080 16GB | 16 GB | Not priced yet | $0.09 | $0.04 | insufficient samples (n=1) |
| RTX 3080 10GB | 10 GB | Not priced yet | $0.05 | $0.04 | insufficient samples (n=10) |
| RTX 6000 Ada 48GB | 48 GB | Not priced yet | — | — | insufficient samples (n=6) |
| RTX 6000 Blackwell 96GB | 96 GB | Not priced yet | — | — | insufficient samples (n=28) |
Data and sources
Data updated: 2026-10-21 (0 days ago)
What the evidence grades mean
- A — official documentation or model card
- B — independent measurement with stated conditions
- C — single-source measurement, not independently verified
- D — derived by calculation; the basis must be stated
Upstream attribution: Nodegrove, LLM VRAM dataset (nodegrove.io), CC BY 4.0. https://doi.org/10.5281/zenodo.22966137
Common questions
Can an RTX 4090 run Qwen3.8-27B at 64K context?
Not with much headroom. At Q4_K_M the weights are 17.8 GB, the KV cache at 64K adds about 4.3 GB (about 0.066 GB per 1K tokens, from the architecture formula and a llama.cpp measurement), and overhead is about 2 GB. That totals about 24.1 GB, right at the 24 GB limit. For 2 GB of headroom, keep the context under about 33K tokens, or use a smaller quant.
Can a 10 GB RTX 3080 run Qwen3.8-27B?
Not by our formula. Even the 2-bit UD-Q2_K_XL weights are 10.7 GB, which already exceeds 10 GB before KV and overhead. Community reports show it running with llama.cpp CPU offload, which our estimate does not count. A 16 GB RTX 4080 fits at short context.
Which GPU does Flux.2 Klein 9B need?
The FP8 version is about 15 GB, which fits a 24 GB RTX 4090 with room to spare (evidence A, official notes). The FP16 version needs about 29 GB.
How much does an RTX 4090 cost to rent here?
Our listed rate is $0.25 per GPU-hour instant and $0.15 queued. We are pre-launch: waitlist only, with no live capacity yet.
Use it from your AI assistant
Connect the SpareHive MCP endpoint and ask fit and price questions directly.
Need a card for a project?
Join the waitlist and describe your workload. We review each request by hand.
Have an idle GPU?
Register your card. We will contact you when a matching request comes in.