RECOMMENDED
Change nothing. This configuration fits comfortably as it is; no upgrade is suggested without a speed target.
Ruled out: Spending on memory or bandwidth buys nothing here: neither is the constraint at this size and context.
Local LLM VRAM & speed calculator
WILL IT RUN ON YOUR MACHINE?
Pick a model and a GPU or a Mac. Get a straight yes or no on whether it fits in memory, an estimated speed in tokens per second, and the one change that fixes it when it does not.
Memory fit is calculated from published specifications and files; speed is an estimate. Neither is a test on your machine. How to read the answer
- Models catalogued
- 327
- GPUs and Macs
- 135
- Models with a format to size
- 320
- Snapshot date
- 2026-10-03
llmb run qwen3-8b --device rtx-4090 --quant Q4_K_M --ctx 8192
weights · Q4_K_M5.03 GB
kv cache · 8,192 ctx1.21 GB
runtime reserve0.80 GB
device memory · RTX 409024.0 GB
memory used7.04 GB / 24.0 GB · 29.3%
It fits
Qwen3-8B · Q4_K_M · RTX 4090
- Tokens / second, est.
- ~116
- Memory needed
- 7.04 GB
- Spare capacity
- 17.0 GB
Calculated, not measured on your machine.Run your own →
Start here
Two ways in
Begin with the machine you own or with the model you want. Either one ends on a page that answers it.
Start from your hardware
I have this machine — what can I run?
Pick your GPU or computer. You will choose its memory configuration next.
Reads the graphics name your browser shares. Nothing is sent anywhere.
Popular:
135 devicesBrowse all →
Start from a model
I want to run this model — what do I need?
Pick a model to see the memory each format needs and which devices hold it.
Popular:
320 models with a format to sizeBrowse all →
Calculator
Check one exact combination
Pick a model, a device, a format and a context length. Every input names its source.
Result updated. It fits in device memory, 29 percent of device memory.
Qwen3-8B · Q4_K_M · NVIDIA GeForce RTX 4090 · 8,192 ctx · offload
High confidencewhy?
The model file's size is published, byte for byte. The runtime's own reserve is an estimate, but it is small against the room left here. Speed is always an estimate, never a measurement on this machine.
89% of the bytes from published sources.
AnswerunvalidatedWeightspublished dataArchitecturepublished dataIT FITS
Fits in device memory: 7.04 GB needed of 24.0 GB of device memory.
Sized for: NVIDIA GeForce RTX 4090, 24 GB, 1008 GB/s
29.3%of device memory · comfortable
- tokens per second
- ~116
- time to first token · floor 0.75s
- ~7.8s
- spare capacity
- 17.0 GB
Faster than you read
for a 8,192-token prompt, one request, no queue
7.04 GB needed of 24.0 GB
Where the memory goes
- Weights · Q4_K_M
- 5.03 GB
- KV cache · 8,192 ctx
- 1.21 GB
- Runtime reserve
- 0.80 GB
- Needed
- 7.04 GB of 24.0 GB published capacity
What to do
NOTHING TO FIXVerdictestimate
17.0 GB spare and ~116 tok/s — above the 30 tok/s this site treats as faster than reading.
Weakest link: the 0.80 GB runtime reserve, which is an uncalibrated prior rather than a measurement
What this answer covers
- File
- a published file with these bytes exists
- Runtime
- not tested for this model and machine — fitting in memory does not guarantee it loads.
- Speed
- estimate, not a measurement
Next step
- Checking before you buy? Where to find the NVIDIA GeForce RTX 4090: Amazon ↗ · eBay (new and used) ↗
- No card at hand? Rent a GPU that runs Qwen3-8B, priced live
Store and cloud links may pay us a commission. They never decide the answer above — the memory arithmetic does.
Snapshot ZOAGzSDaZvVT · catalogue 2026-10-03Open the pair page →
What the labels mean
- published data
- Read from a published source — a file's byte count, a model's configuration or a manufacturer's specification — or exact arithmetic on such values. Not a measurement on a machine.
- estimate
- Calculated from sourced inputs by a stated method; an estimate, not a measurement.
- unvalidated
- Rests on at least one assumption that has not been measured, such as the runtime reserve.
- reconstructed size
- Weight size reconstructed from the pinned architecture, because no published file exists.
- size range
- Only a lower and an upper bound are claimed for this weight size.
Under the hood
How this answer was built
Open what you need. The answer above does not depend on reading any of it.
Memory against context lengthHow the requirement grows as the conversation gets longer.
Sampled values
| Context | Memory | Within 24.0 GB |
|---|---|---|
| 1,024 tokens | 5.98 GB | yes |
| 2,048 tokens | 6.13 GB | yes |
| 4,096 tokens | 6.43 GB | yes |
| 8,192 tokens | 7.04 GB | yes |
| 16,384 tokens | 8.24 GB | yes |
| 32,768 tokens | 10.7 GB | yes |
| 40,960 tokens | 11.9 GB | yes |
The basis of each memory figureWeights 5.03 GB, cache 1.21 GB, reserve 0.80 GB — and which of them is published data.
| Component | Bytes | Basis |
|---|---|---|
| Weights | 5.03 GB | published data published file |
| KV cache | 1.21 GB | published data exact tensor geometry |
| Runtime reserve | 0.80 GB | unvalidated uncalibrated prior |
GPU power ceilingAt most 450 W while generating — the card, not the whole computer.
At most 450 W while generating, so no more than 10.8 kWh a day running continuously.
Why this is an upper bound
- Accelerator energy ceiling at 24 hours a day (default: continuous); this is not the whole machine's bill.
- 450 W is the TGP the manufacturer publishes, which is the most the board may draw — not what decoding draws. Generation is bandwidth-bound and sits below this limit by an amount no vendor publishes, so every figure here is an upper bound.
- Only the accelerators are counted. The rest of the machine — CPU, memory, drives, fans, and the power supply's own conversion losses — draws more on top, and none of it is specific to the model you chose.
- No electricity price was given, so no bill is shown. Rates differ several-fold between countries and there is no default worth quoting.
- No grid carbon intensity was given, so no emissions figure is shown. Intensity varies by more than tenfold between regions and by the hour within one, so a single global average would be a worse answer than none.
Evidence and limitationsHow the file size was obtained, and what was withheld.
Published GGUF file size at a pinned revision.
- No decode fit exists for CUDA with Q4_K_M; the Q4_0 fit on the same backend is used instead.
- Fitted utilisation varies by 1.2% across the three formats calibrated on Metal, which is the measured size of this substitution.
- Calibration sample count is 2 (<5); this calibrated estimate may not generalize.
- Utilisation is solved from one published benchmark on one device and backend. It is applied to other hardware as a first-order extrapolation, not as a measurement of that hardware.
- The compute roof uses the dense matrix-unit peak the vendor publishes, which is the unit the runtime dispatches prefill to. Sparsity-doubled figures are never used.
- Tokenizer time, graph build and prompt-shape effects are outside this equation.
- Published device capacity is a hardware ceiling; display, driver, and runtime allocations can reduce free memory.
Popular answers
What people check most
The pairings people search for, the most downloaded models on the most common cards, every memory size, and card by card.
Can it run?
- granite-4.2-30b on the Ryzen AI Max+ 395
- GLM-5.2 on the Instinct MI355X
- Qwen3-30B-A3B on the Radeon RX 7600
- Qwen3-VL-32B-Instruct on the Apple M5 Pro
- Qwen3-8B on the Radeon RX 9070 XT
- Qwen3.8-Flash-Next on the Ryzen AI Max+ 395
All 18 pairingsFewer
- Qwen3.6-35B-A3B on the RTX 5090
- Qwen3.8-Flash-Next on the DGX Spark
- Qwen3-Next-80B-A3B-Instruct on the Apple M5 Pro
- Qwen3-VL-8B-Instruct on the RTX 3060 12GB
- Qwen3-VL-8B-Instruct on the RTX 5070
- gemma-4-26B-A4B-it on the RTX 3060 12GB
- gemma-4-26B-A4B-it on the RTX 5070
- Qwen3-8B on the RTX 4060
- gemma-4-31B-it on the RTX 4060
- gemma-4-31B-it on the RTX 3050 8GB
- Qwen3.5-9B on the RTX 3050 8GB
- Qwen3.5-9B on the RTX 5060
Best models for your VRAM
What a card runs
Which card to buy →Every card →
Reading the result
How to read the answer
Three questions, in the order the answer gives them. The terms it uses are explained beside them.
Does it fit?
A model has to sit in the graphics card’s own memory to run properly. Three things go in there — the model file, the memory the conversation needs, and what the software reserves for itself — and the answer is whether all three fit in what the card has.
How fast?
Writing each word means reading the whole model out of memory again, so speed is usually decided by how fast the card’s memory is rather than by how powerful the chip is. That is why an older card with fast memory can beat a newer one with slow memory.
What would fix it?
One change, with the number attached — and what will not help, which matters just as much. A faster card does nothing if what you ran out of was memory.
Terms and questions
How much VRAM do I need to run a local LLM?
Enough for three things at once: the model file at the format you choose, the memory the conversation needs, and what the software reserves for itself. As a worked example from this catalogue: Qwen3-8B at Q4_K_M with 8,192 tokens of context needs 7.04 GB, against the 24.0 GB a RTX 4090 has.
What is quantization, and which one should I pick?
Storing the model with less precision so the file gets smaller. Q4_K_M is about a quarter the size of the original and is what most people download; Q8_0 is about half; FP16 is the full thing. Smaller fits on cheaper hardware and answers slightly worse. The usual choice is the largest format that still fits with room for the conversation.
What does context mean?
How much of the conversation the model can see at once, counted in tokens — roughly three quarters of a word each. Longer context needs more memory, and on some models a great deal more.
Why does a long conversation run out of memory?
Because of the KV cache. The memory holding that conversation. It grows as you type, and it is usually what runs out first on a long chat — which is why a model that loaded fine can fail an hour later. Each model-and-device page states the longest context that still fits.
Can I run a model that is bigger than my GPU’s memory?
Often, by offloading. Keeping part of the model in ordinary system memory when it will not all fit on the card. It runs, but slowly: system memory is typically ten to twenty times slower than a graphics card's own.
How accurate is the speed estimate?
It is an estimate, not a test on your machine. The measured error, worst cases included, is published on the accuracy page.
Evidence
Nothing is made up
- 320 models
Read from the publisher’s file
Each model’s architecture is read from the file its own publisher released, at a fixed version we record. Where a figure was never published, this site says so instead of filling the gap.
- Three tiers
The file, or rebuilt from it
When the exact file exists, its published size is used byte for byte. When it does not, the size is rebuilt from the architecture — and the answer tells you which of the two you are looking at, every time. Neither is a measurement taken on a running machine, and the site does not call them one.
- Scorecards / per corpus
Our error is published
Including the cases this engine handles worst, on the same page as the ones it handles best.
See the scorecard →
The same answers, by API
A key is issued the moment you ask for one. 200 calls a day, and a page that shows exactly what you have used.
Put this on your product page
One script tag answers “which models does this card run?” on your own site, and where a figure is not published the widget says so instead of estimating one.