Find the machine for your model. Sized by memory.
41 model entries. Compare GPU memory figures and their sources before choosing a machine configuration.
22 of 28 match the listed memory of one RTX 5090, from $7,880.
By machine.
13 more are sized on request. Search one to ask.
1× RTX 5090
32 GB · $7,880 · cheapest fit for 22Configure & buyModels where 1× RTX 5090 is the cheapest fit
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
qwen36-35b · Profile runtime: ollama · Serve: 30 GB (4-bit (Q4_K_M))
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Format
- Q4_K_M
- Parameters
- 36B
Catalog threshold. Configuration guidance, not measured peak memory. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
vision-llava-34b · Profile runtime: ollama · Serve: 24 GB (Runtime default) · System RAM: 48 GB
Catalog threshold. Configuration guidance, not measured peak memory. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
vision-llava-34b · Profile runtime: ollama · QLoRA: 26 GB (QLoRA)
Planning estimate. Not a B3IQ measurement. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
vision-llava-34b · Profile runtime: ollama · LoRA: 102 GB (LoRA)
Planning estimate. Not a B3IQ measurement. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Fine-tune
- 26 GB · QLoRA
- Parameters
- 34B
Catalog threshold. Configuration guidance, not measured peak memory. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
deepseek-r1-32b · Profile runtime: ollama · Serve: 32 GB (Runtime default) · System RAM: 64 GB
Catalog threshold. Configuration guidance, not measured peak memory. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
deepseek-r1-32b · Profile runtime: ollama · QLoRA: 24 GB (QLoRA)
Planning estimate. Not a B3IQ measurement. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
deepseek-r1-32b · Profile runtime: ollama · LoRA: 96 GB (LoRA)
Planning estimate. Not a B3IQ measurement. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Fine-tune
- 24 GB · QLoRA
- Parameters
- 32B
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
gemma4-31b · Profile runtime: ollama · Serve: 27 GB (4-bit (Q4_K_M))
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Format
- Q4_K_M
- Parameters
- 30.7B
External report. Not independently verified on B3IQ. The profile runtime may differ from the reported setup. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment. View source ↗
Read Qwen3 Coder 30B requirements and sourcesSources and sizing assumptions
qwen3-coder-30b · Profile runtime: ollama · Serve: 32 GB (Runtime default) · System RAM: 64 GB
Catalog threshold. Configuration guidance, not measured peak memory. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
qwen3-coder-30b · Profile runtime: ollama · Serve: 20 GB (4-bit (Q4_K_M)) · System RAM: 32 GB
External report. Not independently verified on B3IQ. The profile runtime may differ from the reported setup. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment. View source ↗
qwen3-coder-30b · Profile runtime: ollama · QLoRA: 23 GB (QLoRA)
Planning estimate. Not a B3IQ measurement. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
qwen3-coder-30b · Profile runtime: ollama · LoRA: 90 GB (LoRA)
Planning estimate. Not a B3IQ measurement. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Other precisions
- Runtime default 32 GB
- Format
- Q4_K_M
- Fine-tune
- 23 GB · QLoRA
- Reported elsewhere
- 30–100 tok/s · 7 reports
- Parameters
- 30B
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
granite42-30b · Profile runtime: ollama · Serve: 23 GB (4-bit (Q4_K_M))
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Format
- Q4_K_M
- Parameters
- 29.3B
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
qwen38-27b · Profile runtime: ollama · Serve: 23 GB (4-bit (Q4_K_M))
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Format
- Q4_K_M
- Parameters
- 27.3B
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
gemma4-26b-a4b · Profile runtime: ollama · Serve: 23 GB (4-bit (Q4_K_M))
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Format
- Q4_K_M
- Parameters
- 25.8B
2× Workstation RTX 5090
64 GB · $19,541 · cheapest fit for 2Configure & buyModels where 2× Workstation RTX 5090 is the cheapest fit
External report. Not independently verified on B3IQ. The profile runtime may differ from the reported setup. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment. View source ↗
Read Llama 3.3 70B requirements and sourcesSources and sizing assumptions
max-70b · Profile runtime: ollama · Serve: 80 GB (Runtime default) · System RAM: 128 GB
Catalog threshold. Configuration guidance, not measured peak memory. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
max-70b · Profile runtime: ollama · Serve: 42 GB (4-bit (Q4_K_M)) · System RAM: 64 GB
External report. Not independently verified on B3IQ. The profile runtime may differ from the reported setup. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment. View source ↗
max-70b · Profile runtime: ollama · Serve: 84 GB (8-bit (Q6_K_L)) · System RAM: 96 GB
External report. Not independently verified on B3IQ. The profile runtime may differ from the reported setup. Sizing note: VRAM includes 20% over the weights for context and runtime. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment. View source ↗
max-70b · Profile runtime: ollama · Serve: 168 GB (16-bit) · System RAM: 192 GB
Model-card-derived planning figure. The row does not record a source URL; not a B3IQ measurement. Sizing note: VRAM includes 20% over the weights for context and runtime. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
max-70b · Profile runtime: ollama · QLoRA: 53 GB (QLoRA)
Planning estimate. Not a B3IQ measurement. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
max-70b · Profile runtime: ollama · LoRA: 210 GB (LoRA)
Planning estimate. Not a B3IQ measurement. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Other precisions
- Runtime default 80 GB, 8-bit (Q6_K_L) 84 GB, 16-bit 168 GB
- Format
- Q4_K_M
- Fine-tune
- 53 GB · QLoRA
- Reported elsewhere
- 15–18 tok/s · 10 reports
- Parameters
- 70B
Catalog threshold. Configuration guidance, not measured peak memory. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
video-hunyuan-xinference · Profile runtime: xinference · Serve: 48 GB (Runtime default) · System RAM: 96 GB
Catalog threshold. Configuration guidance, not measured peak memory. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Parameters
- 13B
4× RTX PRO 6000 Server Edition
384 GB · $99,971 · cheapest fit for 4Configure & buyModels where 4× RTX PRO 6000 Server Edition is the cheapest fit
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
qwen3-coder-480b · Profile runtime: ollama · Serve: 364 GB (4-bit (Q4_K_M))
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Format
- Q4_K_M
- Parameters
- 480.2B
External report. Not independently verified on B3IQ. The profile runtime may differ from the reported setup. Sizing note: VRAM includes 20% over the weights for context and runtime. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment. View source ↗
Read gpt-oss 120B requirements and sourcesSources and sizing assumptions
gpt-oss-120b · Profile runtime: ollama · Serve: 80 GB (Runtime default) · System RAM: 128 GB
Catalog threshold. Configuration guidance, not measured peak memory. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
gpt-oss-120b · Profile runtime: ollama · Serve: 72 GB (4-bit (MXFP4)) · System RAM: 64 GB
External report. Not independently verified on B3IQ. The profile runtime may differ from the reported setup. Sizing note: VRAM includes 20% over the weights for context and runtime. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment. View source ↗
gpt-oss-120b · Profile runtime: ollama · QLoRA: 90 GB (QLoRA)
Planning estimate. Not a B3IQ measurement. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
gpt-oss-120b · Profile runtime: ollama · LoRA: 360 GB (LoRA)
Planning estimate. Not a B3IQ measurement. No row-specific sizing note is recorded. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Other precisions
- Runtime default 80 GB
- Format
- MXFP4
- Fine-tune
- 90 GB · QLoRA
- Reported elsewhere
- 70–145 tok/s · 11 reports
- Parameters
- 120B
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
llama4-scout · Profile runtime: ollama · Serve: 85 GB (4-bit (Q4_K_M))
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Format
- Q4_K_M
- Parameters
- 108.6B
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
Sources and sizing assumptions
qwen3-coder-next · Profile runtime: ollama · Serve: 65 GB (4-bit (Q4_K_M))
Planning estimate. Not a B3IQ measurement. Sizing note: Weight-size floor plus 25% allowance; not hardware-qualified. Context, image resolution, frames and concurrency can require more. Test date, measured context length, concurrency, runtime version and test hardware are not recorded for this row. Confirm these before deployment.
- Format
- Q4_K_M
- Parameters
- 79.7B
These are catalog sizing thresholds, planning estimates and external reports. Context, concurrency and runtime overhead can increase memory use. A memory match does not establish a tested deployment, throughput or live capacity. “Runtime default” means the sizing row does not name a precision. Fine-tune figures use QLoRA where recorded, otherwise LoRA.
