Stefan Krassin

CEO

[Stefan Krassin @stefan_krassin]

Revolut’s AI ops lead: out-of-the-box inference is terrible. Tuning cut their time to first token by 88% on the same GPUs.

We measured that gap on 9 open models: standard vLLM vs. tuned by yasp. Up to 2.3× more tokens/s, accuracy-checked.

Benchmarks: https://t.co/It3uxlDawV

[Stefan Krassin @stefan_krassin]

Inference costs keep growing, and most teams depend on one hardware vendor.

So we tested AWS #Trainium3: our agents ran Qwen3.5-35B-A3B at 11,400 tok/s on one instance, 512 concurrent requests.

That’s 81% of an 8-GPU server’s throughput, same workload.

More models soon.

[Stefan Krassin @stefan_krassin]

Same Trainium2 chip. Same model. Same price per hour.

Up to 83% more tokens per second. Up to 45% lower cost per token.

Before you order more hardware, check what the chips you already pay for can actually do.

Big thanks to @awscloud Trainium Team and @PyTorch #TorchNeuron

[Reza Rahimi @reza_rahimi_ca]

What I care about here: more throughput from the same chip AND faster responses per user.

Qwen3.5-35B-A3B on one Trainium2 chip: 427 tok/s total, ~53 tok/s per user, +83% yasp optimized vs. initial port.

Big thanks to @awscloud Trainium Team and @PyTorch #TorchNeuron.

[Stefan Krassin @stefan_krassin]

Full story: https://t.co/N3ReYlGUyd

[Stefan Krassin @stefan_krassin]

AI hardware is often picked by where a model was tuned first, not by price-performance.

With @awscloud we took Qwen3.5 to its best on Trainium2 in days:
→ up to 83% more tokens/s
→ up to 45% lower cost per token
→ same outputs

Same chip. Far more tokens for the money.