One model, two PCs. You be the judge.

      
      DeepSeek-R1-Distill-Qwen-7B
      · the model on both machines

The same model, on two identical RTX 4090s. PC A is pressed by Trapetum to 4-bit (5.5 GB); PC B runs it uncompressed in fp16 (15.2 GB). Same weights, same card, same prompt, the only difference is the compression. Send one prompt to both and read the tokens per second yourself.

PC A, Trapetum 4-bit

5.5 GB

4-bit weights · pure-Rust runtime · RTX 4090

PC B, fp16 uncompressed

15.2 GB

stock PyTorch / transformers · RTX 4090

Both PCs are rented RTX 4090s (RunPod community cloud), specs pinned. During your test each GPU serves you alone, batch-1, exactly like running these two machines at home.

Sessions used today: 0/12 · type a prompt and hit Send to start a 5-minute test.

Your machines · time left 5:00

DeepSeek-R1-Distill-Qwen-7B the model on both machines

PC A, Trapetum 4-bit · 5.5 GB

, tok/s

Waiting for your prompt…

PC B, fp16 uncompressed · 15.2 GB

, tok/s

Waiting for your prompt…

Why these numbers are real

During your session, both GPUs serve only you. A page-wide lock hands the two machines to one tester at a time, batch-1, exactly what you would see at home with these two PCs. Other visitors affect only the wait in line, never your measured speed. The limits (five minutes per test, twelve tests per day) exist so everyone gets the machines to themselves.

Same model, same card. Identical DeepSeek-R1-Distill-Qwen-7B weights on two identical RTX 4090s. The only variable is Trapetum's 4-bit compression on PC A.
The clock is honest. Each panel's tok/s is measured only between the first and last streamed chunk, network latency to the pod is excluded from the rate.
Nothing is cached. Both PCs decode every prompt fresh; nothing is stored between sessions.

Curious how the 4-bit weights get there, and why decode is bandwidth-bound? The full method and raw numbers live on the benchmarks page. New to Trapetum? Read what Trapetum does, in plain language, see the models that fit a 24 GB card, or install Trapetum on your own machine.

Both PCs are rented RTX 4090s (RunPod community cloud), specs pinned; nothing is cached between sessions. Live figures are measured on this specific hardware and configuration and are provided for information only, not as a performance guarantee. Terms.