# Package the model and get a deploy to come up

> Package the model and get a deploy to come up: 5 companies fully do it, 6 with a limit, 24 carry the tag. Every listing in GPU & serverless compute with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/gpu-compute.package-and-deploy/
- Markdown: https://www.anchorterminal.com/niches/gpu-compute.package-and-deploy/index.md (~1,800 tokens)
- Slim: https://www.anchorterminal.com/niches/gpu-compute.package-and-deploy/index.min.md (~280 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/gpu-compute.package-and-deploy/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 5 | 6 | 24 | 8 | 6 |

## What counts as done

A build and deploy from the CLI or API completes, the worker reports ready with the engine loaded, logs of a failed start are readable, and no secret is printed in the deploy output.

Measured, for our own tests: the build and deploy complete within 20 minutes.

How agents ask for it: deploy this repo as a GPU function; push my Cog model to Replicate; build a vLLM worker for this model and deploy it

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Baseten (Baseten Labs, Inc.)**: baseten model push builds and deploys from config.yaml; the Startup time card shows Failed and Stalled (30 minutes) outcomes with the phase reached and links to logs; team keys authenticate CI without a browser.
- **Cerebrium (Cerebrium Inc.)**: cerebrium deploy runs headless with CEREBRIUM_SERVICE_ACCOUNT_TOKEN, builds are cached and billed at $0.08 a minute, a GPU outside the plan is rejected at deploy time, and build logs are readable through the REST API.
- **Cloud Run GPUs (Google LLC)**: gcloud run deploy --source builds on an e2-highcpu-8 and deploys with --gpu 1 --gpu-type; a startup probe can be lengthened for slow model loads and logs are in Cloud Logging. A service account key or workload identity authenticates CI without a browser.
- **Koyeb (Koyeb SAS)**: Services build from git or a registry image and the CLI authenticates headless with a token; dry_run validates a definition before deploying and instance statuses (Starting, Unhealthy, Error) with messages are exposed for failed starts.
- **Northflank (Northflank Ltd)**: Services and jobs build from a linked git repository or deploy a registry image, API tokens come from roles and work headless, and build and job logs are API endpoints.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Beam (Smartshare, Inc.)**: beam deploy --json with BEAM_TOKEN works headless from CI and the platform API streams logs. The founder-reported issue 1828 showed --format json printing the workspace token into CI logs until PR 1886 (8 September 2026), so older CLIs still leak.
- **Hugging Face Inference Endpoints (Hugging Face, Inc.)**: A custom container or an engine config is deployed through the UI, API or Python client with runtime logs available, but there is no build step: the image must already exist in a registry, and the FAQ covers endpoints stuck at initializing.
- **Modal (Modal Labs, Inc.)**: modal deploy builds the image and deploys from a token in the environment; crash-looping containers appear in the dashboard. A deployed Function that crashes on start is retried indefinitely with backoff rather than stopped, and there is no REST deploy API.
- **Replicate Deployments (Replicate, LLC)**: cog push r8.im/user/model builds and publishes with REPLICATE_API_TOKEN set headless, but a push of a granite-speech model failed with unknown blob for two months (acc-41) and Cog 0.23.1 flooded receivers with full-log webhooks that restarted the pod mid-prediction (acc-42).
- **Runpod (Runpod, Inc.)**: Templates and endpoints are created through REST v2 or the CLI, and rp login takes RUNPOD_API_KEY headless since 2.1.0. A startup OOM in the vLLM worker exits 1 and restarts on the same footprint with no response-level sign of the eager relaunch (acc-09).
- **Verda (DataCrunch Oy)**: Containers deploy from any registry through the REST API with Docker config.json or provider auth, but there is no build step: the image must exist first, and the ECR note warns pulls fail after 12 hours without the token permission.

- not this job: CoreWeave (CoreWeave, Inc.): No build or push step; customers bring container images to Kubernetes or point Dedicated Inference at weights in object storage.
- not this job: Crusoe Cloud (Crusoe Technologies LLC): No image build or push; Crusoe Container Registry stores images for CMK and VMs take custom images, but nothing deploys a worker from a repo.
- not this job: GPUsMarket (Tyga.Cloud Ltd): No custom image build or push; instance types are the host's Ollama, a docker container with SSH, or a shell.
- not this job: Hyperbolic (Hyperbolic Labs, Inc.): No image build, push or deploy step; the product is a raw VM with Docker installed.
- not this job: Lambda Cloud (Lambda, Inc.): No image build, push or deploy; the product is a VM with Lambda Stack preinstalled.
- not this job: Massed Compute (Massed Compute, Inc.): No image build, push or deploy; images are a fixed list of VM images.
- not this job: Nebius AI Cloud (Nebius B.V.): Serverless AI takes a container image from a registry; there is no build or push step on the platform.
- not this job: Thunder Compute (Thunder GPU, Inc.): Sandbox images can be prepared from a Dockerfile, but nothing deploys a worker or endpoint from them; instances are VMs with templates.
- not this job: Together GPU Clusters (Together Computer, Inc.): No image build or push; workloads are Kubernetes manifests or Slurm scripts the user brings.
- not this job: Vast.ai (Vast.ai Inc.): Templates reference an existing Docker image; there is no build or push step on the platform.

## What stops agents finishing

- engine defaults tuned for one GPU list OOM on a card not in the list, and the worker restarts on the same footprint
- image push fails with an opaque registry error and the only workaround is retrying
- deploy output in JSON mode contains the bearer token, so CI logs leak it
- CLI login opens a browser, so a headless agent needs a separate service token

## Around it

Before: open-weights.fine-tune-base, agent-harnesses.fix-failing-ci. After: gpu-compute.serverless-model-endpoint.

## Where this comes from

Weight 8 of 100 in GPU & serverless compute: 6 of 48 accounts are deploy-time failures: startup OOM with a default setting, a push that fails with unknown blob for two months, a client that ignores Retry-After, a deploy log that prints the workspace token.

As data: https://www.anchorterminal.com/api/v1/use-cases/gpu-compute.package-and-deploy.json, or the MCP tool get_use_case with id gpu-compute.package-and-deploy. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
