One model, two RTX 4090s. You be the judge.

The same DeepSeek-R1-Distill-Qwen-7B on two identical RTX 4090s: one pressed to 4-bit by Trapetum (5.5 GB), one left at full fp16 (15.2 GB). Same prompt, both cards. This is a recording of a real run, replayed at each side's true measured speed.

recorded live on two RTX 4090s, July 8 2026

PC A, Trapetum 4-bit

113.1 tok/s
5.5 GB on disk, measured decode speed

PC B, fp16 uncompressed

37.4 tok/s
15.2 GB on disk, measured decode speed
DeepSeek-R1-Distill-Qwen-7Bthe model on both machines
PC A, Trapetum 4-bit · 5.5 GB
tok/s
Press Play to watch both cards answer at their real speed.
PC B, fp16 uncompressed · 15.2 GB
tok/s
Same prompt, same GPU model.
Explain quantization to a 12-year-old.

Why these numbers are real

Both consoles replay the actual token stream each RTX 4090 produced, at each side's measured decode speed. Trapetum runs the 4-bit model in its pure-Rust fused kernel; PC B runs the untouched fp16 weights the way a home user would. Same model, same card, a third of the memory, about 3.0x the speed. Live decode speed varies run to run (fp16 especially), and often reaches 4x.

See the full per-GPU benchmarks, watch the 671B on a single 4090, or follow the Road to x100. Numbers here are decode tokens per second, measured on the endpoints, prompt "Explain quantization to a 12-year-old.".

And the full 671B, inverted: 2.46 tok/s on one machine

Compression is not only about speed. The same 4-bit codebook idea lets the largest open-source model, DeepSeek-R1 671B, run on a single commodity node: the routed experts stay resident in host RAM and the CPU decodes them in place, while the GPU holds only attention. Replayed token by token at its measured steady-state speed of 407 ms/token.

2.46 tok/s
steady state (x10.2 vs disk-offload)
671B
parameters (256-expert MoE)
20.9 GB
VRAM: attention + dense only
350 GB
4-bit experts, resident in RAM
DeepSeek-R1 671Bcommodity node · CPU-experts · greedy
The capital of France is
Press Play to watch the 671B think at real speed.
token 0/59elapsed 0:00last token -speed -

The full run takes about 24 seconds at real speed. Decode is bound by host memory bandwidth now, not disk: the experts are decoded in place on the CPU at ~67 GB/s aggregate. See the benchmarks and the Road to x100.

671B at 2.46 tok/s: with vs without Trapetum

Same model (DeepSeek-R1 671B, 1.34 TB in bf16). Same peak speed. Only the hardware and the cost change.

PEAK SPEED HELD AT 2.46 tok/s ON BOTH CONFIGS
Without Trapetum
standard stack: weights held in GPU VRAM
PEAK SPEED 2.46 tok/s
Weights in VRAM671 GB FP8 / 1.34 TB bf16
GPUs8× H200 or 16× H100
Servers1 to 2 DGX-class nodes
HARDWARE
~$400-500K 8× H200 DGX
Cloud ~$30-50 / hr
With Trapetum
CPU-experts inversion, lossless 4-bit
PEAK SPEED 2.46 tok/s
Experts in host RAM350 GB (4-bit)
VRAM used20.9 GB · 1× used 3090 / L40S
Serverone node, 64-core + 512 GB RAM
HARDWARE
~$6-9K refurb DDR4 · $28K all-new
Cloud ~$4-5 / hr
Same 2.46 tok/s → a ~$6-9K box vs a ~$450K cluster (~50× cheaper)

Without the inversion a single GPU tops out at 0.33 tok/s (naive offload, measured), so the weights must go in VRAM: a cluster. The job needs 350 GB at only ~67 GB/s, which refurbished DDR4 delivers for ~$8K; all-new DDR5 (512 GB ~$15K) makes it ~$28K. Street prices, July 2026.