Keep your current machine
If the model and context fit your actual setup, test that first. Do not buy a larger GPU just because a parameter count sounds large.
See the existing-machine checks →Buying guide
Check the model before you buy the machine
Start with the workload and context you actually need. Then choose the smallest change that supports it: keep your current computer, add a GPU to a compatible desktop, buy a complete system, or compare ownership with occasional external compute.
Choose a path
These are decision prompts, not a ranking. Fit, context, runtime support, workload frequency, and what you already own all change the answer.
Add a GPU to a desktop you own
Compare the added VRAM with your target model, then verify the desktop has room, a compatible slot, adequate power delivery, and cooling. A GPU MSRP is card-only.
See the GPU upgrade checks →Buy a complete system
A system price covers a configured computer rather than only a graphics card. Check the exact memory and storage configuration and leave room for the operating system and other apps.
See complete-system checks →Use external compute for occasional work
When work is infrequent, compare the system you would actually buy against your own current rental rate and usage estimate. The calculator does not prefill changing cloud prices.
Compare costs in the calculator →Worked checks
Same checker, real hardware and model records
Every result below is computed from the committed dataset with llama.cpp, batch 1, 8K context, fp16 KV cache, and the published engine assumptions. A fit verdict is not a speed benchmark or a guarantee that your drivers, app, and other memory use will match. Purchase search links appear only beside clean fits with a non-painful tier; tight and no-fit examples stay linked to the checker. Read the methodology and open the exact checker state before deciding.
Keep what you have, if it passes your check
An RTX 4060 can be enough for a smaller target and still have a hard limit on a much larger one.
Keep the machine you already own
NVIDIA GeForce RTX 4060
A concrete check before spending: an RTX 4060 with Qwen3-8B at 8K context.
Qwen3-8B · 8K context
Q4_K_M · llama.cpp · highest-quality quant with a clean fit in this check
Fits · Grade BInteractive speed tier
No CPU offload in this result
Dataset MSRP: $299 (card only; as of 2026-07-08). Historical source value, not a current offer or live market price. It excludes the desktop, power supply, cooling, and any other parts you may need.
Sources: GPU specification and recorded MSRP · model source
A real limit on the same GPU
NVIDIA GeForce RTX 4060
Check a much larger model on the same card. If no quant fits cleanly, this example evaluates the smallest listed artifact.
gpt-oss-120b · 8K context
MXFP4 · llama.cpp · smallest listed quant; no clean fit was found
Won't fit · Grade D
No CPU offload in this result
Beta caveat
- Hybrid attention — KV computed as full-cache upper bound
Dataset MSRP: $299 (card only; as of 2026-07-08). Historical source value, not a current offer or live market price. It excludes the desktop, power supply, cooling, and any other parts you may need.
Sources: GPU specification and recorded MSRP · model source
Add a GPU only for the headroom you need
Compare two model sizes on the same 16GB card; the computed verdict and offload labels show the difference.
Add a GPU for a desktop you already own
NVIDIA GeForce RTX 5060 Ti 16 GB
Check this 16GB card with Qwen3-14B at 8K before treating it as an upgrade.
Qwen3-14B · 8K context
Q5_K_M · llama.cpp · highest-quality quant with a clean fit in this check
Fits · Grade BInteractive speed tier
No CPU offload in this result
Dataset MSRP: $429 (card only; as of 2026-07-08). Historical source value, not a current offer or live market price. It excludes the desktop, power supply, cooling, and any other parts you may need.
Sources: GPU specification and recorded MSRP · model source
Open this exact check →
Search this GPU on Amazon
Affiliate link — we may earn a commission. Verdicts are computed before any link is attached.
The 16GB card's larger-model boundary
NVIDIA GeForce RTX 5060 Ti 16 GB
See how Qwen3-32B changes the memory requirement on the same card, including any CPU-offload caveat.
Qwen3-32B · 8K context
Q4_K_M · llama.cpp · smallest listed quant; no clean fit was found
Tight · Grade CUsable speed tier
CPU offload used
Beta caveat
- CPU offload — speed varies heavily with layer split
Dataset MSRP: $429 (card only; as of 2026-07-08). Historical source value, not a current offer or live market price. It excludes the desktop, power supply, cooling, and any other parts you may need.
Sources: GPU specification and recorded MSRP · model source
A complete system changes the ownership question
These examples are separate checks, not a head-to-head speed ranking. A listed system price is a dated source value, not a live offer; confirm the current configuration with the vendor.
Complete system: Mac Studio M4 Max 64GB
Mac Studio M4 Max 64GB
A sourced 64GB unified-memory system checked with Qwen3-32B at 8K.
Qwen3-32B · 8K context
Q8_0 · llama.cpp · highest-quality quant with a clean fit in this check
Fits · Grade BUsable speed tier
No CPU offload in this result
Dataset-listed base-system price: $3,499 (as of 2026-07-11). Historical source value, not a current offer or live market price.
Sources: manufacturer system listing · model source
Open this exact check →
Search this system on Amazon
Affiliate link — we may earn a commission. Verdicts are computed before any link is attached.
Complete system: NVIDIA DGX Spark
NVIDIA DGX Spark
A 128GB unified-memory system checked with Llama 3.3 70B at 8K; its speed tier remains beta for this hardware class.
Llama-3.3-70B-Instruct · 8K context
Q8_0 · llama.cpp · highest-quality quant with a clean fit in this check
Fits · Grade APainful speed tier
No CPU offload in this result
Beta caveat
- New hardware class — tier not yet calibrated
Dataset-listed base-system price: $4,699 (as of 2026-07-11). Historical source value, not a current offer or live market price.
Sources: manufacturer system listing · model source
Before you spend
- Run the exact model, quant, runtime, and context you plan to use.
- For a GPU, count the whole desktop path: slot, clearance, power supply, and cooling.
- For unified memory, the OS and other apps share the same pool; total RAM is not all model memory.
- For occasional work, compare using your actual current rate and expected usage rather than a stale price.