Reference

Models — what we serve and what we've measured

Every latency below was measured by us, on the listed GPU, chunk-level (one full policy inference producing an action chunk), p50 over ≥15 runs after warm-up, with the sampling path torch.compile'd unless noted. Full logs and methodology are shared with design partners.

Serving today

Policy Params Chunk H100 p50 L4 p50 Status
SmolVLA (lerobot/smolvla_base) 450 M 50 actions 22.7 ms (p95 22.8) 48.3 ms ✅ Serving
SmolVLA, reduced denoise (steps=1) 450 M 50 actions 12.5 ms 29.4 ms ✅ Opt-in (quality A/B first)

SmolVLA is a flow-matching policy: latency ≈ fixed cost (vision/prefix) + per-denoise-step cost. Measured decomposition — H100: 26.6 ms fixed + 14.0 ms/step (eager); L4: 54.9 + 27.3. Compiling the sampling path (sample_actions, mode="reduce-overhead") is worth 6.8× — this is the core of why our numbers beat naive serving.

Benchmarking / integrating

Policy Params State Blocker
π0 (Physical Intelligence, open weights) ~3.3 B Benchmarking none — same lerobot path as SmolVLA
GR00T N1.7 (NVIDIA) ~3 B Integrating pip lerobot wraps N1.5 only; needs Isaac-GR00T native install (productization in progress)
OpenVLA 7 B On request autoregressive (no chunking) — different latency profile; best for ≤5 Hz replan

Your own checkpoint

Fine-tuned a SmolVLA/π0 variant on your task data? That's the expected case — base checkpoints are demos; your fleet runs your weights. We deploy safetensors checkpoints during onboarding, re-run the latency harness on them, and hand you the measured p50/p95 before anything touches a robot. Same architecture ⇒ same latency as the table above; new architectures get benchmarked first.

Choosing a tier

  • H100 — you have a hard sub-30 ms deadline (high-rate visual servoing, humanoid balance-adjacent stacks): 22.7 ms full quality clears it with margin.
  • L4 — replan at ≤10 Hz (pick-and-place, kitting, most warehouse manipulation): 48.3 ms is far inside your budget at a fraction of the cost.
  • Undecided? We run your policy on both during onboarding and show you the numbers.