localLLMriglocal, on your table

Ready now

All the models are ready.

Pick any, run your prompt, see the answers side by side.

Free LLM Showdown: 6 open models, 10 real prompts →

33 open models ready. Pick any, run your prompt, see the answers side by side.

9 of 9 open models selected. Nine open models is the most one comparison can run. Remove one to add another.

Small & fast
Mid-size
Large & smart
Code
Multilingual
Mixture of experts
Frontier reference

Example

Example. These cards show the layout, not a model answer. Speed is not a measurement. Cost is the catalog list price.

Example

Llama 3.2 3B Instruct

Your prompt's answer appears here.

Speed About 242.4 tok/s on Used PC with RTX 3090 24GB. Estimated, not measured.

Cost List price, 2026-10-15: $0.00043 for a 2,000 in / 1,000 out cap. Not an invoice.

Example

Gemma 3 12B

Your prompt's answer appears here.

Speed Not measured.

Cost List price, 2026-10-15: $0.00025 for a 2,000 in / 1,000 out cap. Not an invoice.

Example

Qwen3 32B

Your prompt's answer appears here.

Speed Not measured.

Cost List price, 2026-10-15: $0.00044 for a 2,000 in / 1,000 out cap. Not an invoice.

Example

Llama 3.3 70B Instruct

Your prompt's answer appears here.

Speed About 17.8 tok/s on Dual RTX 3090, 48 GB. Estimated, not measured.

Cost List price, 2026-10-15: $0.00197 for a 2,000 in / 1,000 out cap. Not an invoice.

  1. 1

    Pick the models

    Every open model is ready. Nine are already selected. Click any chip to add or remove it.

  2. 2

    Run your prompt

    Type a question your business would ask. One prompt is enough. You can change the models and run it again.

  3. 3

    See the answers side by side

    Each answer shows speed and cost. A rented GPU uses a checked hourly rate times hours, labelled as an estimate. A local price is a purchase estimate, not a monthly bill.

Not sure where to start? Pick a use caseAlready know your model? Size your hardware

Hardware estimate

VRAM calculator: which hardware do you need to run your LLM locally?

Give a budget, a model, or both. You get three concrete setups: enough memory or not, an estimated speed, and a price we label as an estimate. Global prices (USD), EUR available.

12 models · 19 setups

Your question

v1

Examples

Common questions

Does Llama 70B fit on an RTX 4090?
Not in full at Q4_K_M. Weights are about 42 GB in this formula, about 46 GB with KV cache and overhead at 8,192 tokens, and a 4090 has 24 GB. Two 24 GB cards, or a Mac with enough unified memory, can at that context. That is still an estimate, not a measured GGUF. At 32,768 tokens with FP16 KV the same pair does not fit.
Are the tokens per second measured?
No. Tokens per second are estimated from memory bandwidth. There is no benchmark source: benchmarkSourceUrl stays null.
Are the prices accurate?
On 16 October 2026 the stored RTX 3090 card is the used median €1,224 and the stored RTX 4090 card is the used median €2,185, both from pcprice.watch. The same cards also show GPU PRIX new prices: 3090 €2,218 and 4090 €3,790. The RTX 5090 card is €6,299 from PriceSquirrel (Proshop.fi, in stock). Build totals use those stored card prices plus a sourced RAM line and rest of system (estimate), and stay unverified. Mac mini M6 16 GB, Mac Studio M5 Max 36 GB and M5 Ultra 96 GB cite the Belgian Apple Store. Other Apple euro amounts are estimates or used figures and say price unverified. Rows still dated 2026-10-14 without a listing say price unverified. EUR includes VAT. USD is an FX estimate.
Does the newsletter store my address?
Only when the newsletter is open. No email is sent. While it is closed, nothing is stored.
All the models are ready · localLLMrig