NVIDIA Brev Launchable

Launchable for NVIDIA Brev

Run the public TeleMemetry benchmark on a provisioned NVIDIA Brev GPU instance
The launchable prepares the instance and starts the local benchmark page automatically
Each run produces downloadable SHA256 evidence bundles for independent review

Launch on NVIDIA Brev(~7 minutes setup - 1 minute per run) View reproduction repo

Billing note: Brev GPU instances can keep billing while they are running. Download your evidence bundle, then delete the Brev environment when you are finished.

Video walkthrough

Deployment, first run, and evidence bundle download - step by step. No audio in this version - apologies. A narrated version is coming soon.

What to expect

Provisioning and setup can take up to 10 minutes on first launch.

Brev provisions a GPU instance

A fresh GPU instance is allocated in your Brev environment. You may need Brev balance for GPU time.

Startup script pulls the benchmark repo

The latest public reproduction harness is cloned from github.com/TeleMemetry/reproduce.

The local benchmark service starts automatically

Dependencies install and the benchmark server starts. This phase takes most of the setup time.

Services healthy - click the benchmark link in Brev

When the service is ready, Brev shows a link to the local benchmark page. Click it to open the interface.

Benchmark runs produce result files and evidence bundles

Each run writes SHA256 receipts, a per-probe scoring log, and a versioned downloadable evidence bundle.

Visual walkthrough

Step-by-step screenshots

Hover any screenshot to zoom. Add your notes in the comment area below each image.

01 Launchable deploy page

L40S (48 GiB), 12 CPUs, 72 GiB RAM on MassedCompute. Hardware and cloud provider are pre-selected. Click Deploy Launchable.

02 Sign in to NVIDIA

NVIDIA account required for Brev billing. Enter your email to sign in or create a free account.

03 Confirm and deploy

Signed in. Brev auto-generates a unique instance name. Confirm the instance type and click Deploy Launchable to start provisioning.

04 GPU provisioning

Brev allocates the L40S instance on MassedCompute. Configure, startup script, and health check run automatically in sequence.

05 All systems ready

All four steps show green checkmarks. GPU provisioned, instance configured, startup script ran, services healthy. Click memory-demo to open the benchmark.

06 Verify + Package

The benchmark app is open. 3,000 turns, 10 fields, 20 episodes are pre-filled. Click Verify + Package to generate evidence artifacts and SHA256 receipts.

07 Verification complete

PASS. Result package is ready in results/latest. All four pipeline stages are green: evidence artifacts, verified results, SHA256 receipts, review bundle.

08 Results summary

3,000 of 3,000 turns verified. 0 failures. 52.96 tokens per turn. 192.79x context-history reduction estimate. Download links for the full evidence bundle appear below.

09 Machine-readable result file

RESULT: PASS. Verified recall 3000/3000, zero output failures, 52.96 average tokens per turn, 192.79x replay reduction. Source commit and launchable version recorded for audit.

10 GPU Environments dashboard

Brev GPU Environments dashboard. Instance is Running, container Built. The memory-demo service link and Jupyter notebook are both accessible. If you see this bottom panel you are paying for your instance - click it to delete it to stop billing.

11 Secure Links

Instance detail view. Port 7860 (benchmark app) and port 8888 (Jupyter) both show Healthy. The benchmark URL is shareable with teammates or auditors.

12 Delete when done

When your run is complete, delete the environment to stop billing. Download your evidence bundle first - data cannot be recovered after deletion.

Before you launch

You may need several dollars of Brev balance for GPU time - check your account before provisioning.
Keep the instance only as long as needed - delete the Brev environment when the benchmark run is complete to stop billing.
Download the evidence bundle before deleting the instance - it will not be recoverable after.
Each benchmark run creates its own versioned evidence bundle filename - you can run multiple times and compare.
The benchmark is a public reproduction harness, not the private production engine.

What this benchmark demonstrates

The public Launchable demonstrates bit-perfect operational recall within the public benchmark scope. It verifies that the probe-answer-score pipeline produces the reported numbers against the public nuPlan telemetry subset, and that the resulting evidence bundle is SHA256-verifiable.

It does not prove production TeleMemetry internals. It does not claim to replace all databases, RAG, compression, or pruning in all systems. The certified registry runs in the evidence library are produced by the private production harness and carry the full artifact chain required for engineering due diligence.

Read the full IP boundary statement →

Ready to run

Launch on NVIDIA Brev

Click the button below to open the NVIDIA Brev deploy page. Provisioning and setup take up to 10 minutes. When the service is healthy, click the link shown in Brev to open the local benchmark interface.

Billing note: when the run is finished, download your evidence bundle and delete the Brev environment. Closing the benchmark page does not stop GPU billing.

Launch on NVIDIA Brev(~7 minutes setup - 1 minute per run) View reproduction repo