Verified edge models, without owning a GPU.
Upload a model, a target chip, and a latency budget. FastCrest Cloud optimizes it on rented GPUs and proves the optimized policy still behaves like the original, then hands back a verified edge artifact and a signed parity cert. No TensorRT wrangling, no GPU bill.
How it works
Upload, prove, ship.
Three steps. Each one writes a verifiable artifact, and the job refuses to hand back a model that fails its parity gate.
Model + chip + budget
Point at a HuggingFace checkpoint or upload your own export, name the target board (Orin Nano to Thor), and set a latency budget.
Cloud GPUs do the work
Distill, quantize, and compile for the chip, then run the parity suite: a distribution test, rollout success-rate, and embodied metrics against the original policy.
Artifact + signed cert
Get a verified edge artifact, a signed parity.cert.json, a human-readable PARITY.md, and a reproducible manifest.
What you get
Every job ships a paper trail.
A FastCrest Cloud export isn't just a faster model. It's a model you can prove is still the model.
Verified edge artifact
ONNX plus a TensorRT engine targeted at your exact board and precision. Drops straight into tether serve or your own runtime.
Signed parity cert
parity.cert.json + PARITY.md
A cryptographically signed record of how the optimized policy compares to the original, on the metrics that matter for control.
Reproducible manifest
manifest.json
SHA-256 of every file in the bundle, so anyone downstream can confirm they have exactly what was certified.
CI gate
Wire the same check into your pipeline. A deploy that regresses parity past your threshold fails the build instead of reaching a robot.
The difference
Most tools make it faster. FastCrest Cloud proves it didn't change.
Quantizing and distilling a policy can silently shift its behavior in ways a latency number never shows. The parity proof catches that.
Distribution test
A two-sample test (MMD) over action chunks detects behavioral drift from a small number of rollouts, with false-positive guarantees.
Rollout success-rate
Task success measured on the optimized policy versus the original, so a cert reflects what the robot actually does, not just tensor closeness.
Embodied metrics
Jerk, completion-time, and motion-energy, because a policy can hit the same success rate while moving in a way you'd never ship.
Pricing
Pay per outcome, not per GPU-hour.
You're buying a verified artifact, not compute time. We cap the GPU with a hard watchdog so a runaway job is our problem, not your invoice.
- Optimize: distill / quantize / compile for the chip
- Prove: parity cert (distribution + rollout + embodied)
- Verified artifact, cert, and manifest to download
- Hard GPU watchdog, billed on outcome
- Blocks a deploy on parity regression
- Your threshold, your metrics
- Runs against each candidate model
- Same cert format as per-export jobs
Cert
What you actually get back.
Every verified job returns a signed parity.cert.json, a human-readable PARITY.md, and the optimized ONNX artifact. The cert is reproducible — anyone with the cert and the original model can re-run the verification.
{
"schema_version": "1.2.0",
"model_ref": "pi05_libero_finetuned_v044",
"target_sku": "jetson_orin_agx_64gb",
"runner": "tether-verify@0.11.2",
"n_episodes": 100,
"success_rate_native": 0.833,
"success_rate_optimized": 0.967,
"mmd_p_value": 0.008,
"distributions_differ": false,
"embodied_metrics": {
"jerk_p95": 2.41,
"motion_energy_mean": 0.182,
"completion_time_s": 4.72
},
"attestation": {
"signed_by": "fastcrest-verify",
"key_id": "fc-prod-2026-05",
"timestamp": "2026-06-01T18:42:11Z",
"signature": "ed25519:9f3a…b71c"
},
"manifest_sha256": "a4b1c83e…7d2f"
}| Metric | Native | Optimized |
|---|---|---|
| Success rate | 0.833 | 0.967 |
| MMD p-value | 0.008 (no significant drift) | |
| Jerk p95 | 2.41 m/s³ | |
| Completion time | 4.72 s | |
Verify the signature yourself: tether verify cert ./parity.cert.json --key fc-prod-2026-05
Request early access
FastCrest Cloud is in early access. Tell us the model and board you'd verify first and we'll get you in.
