How to run Runway Gen-4.5 and Aleph 2.0 through Baseten Dedicated Inference

Short answer: running Runway Gen-4.5 or Aleph 2.0 on Baseten is an enterprise onboarding project, not a matter of calling a publicly documented shared endpoint. You license the model weights directly from Runway, ask Baseten to provision a single-tenant workload plane with reserved GPU capacity, deploy the licensed model, and then connect your application using the endpoint and credentials issued during onboarding. Baseten isolates the model workload and manages the control plane, infrastructure and scaling.

Baseten announced this deployment path on October 14, 2026. The announcement does not publish a Baseten URL, authentication header format or deployment command, so those values must come from your Baseten and Runway account teams rather than being guessed from Runway’s public developer examples.

#	Preview	Product	Price
1		HPE NVIDIA Tesla V100 32GB HBM2 PCIe 3.0 x16 Passive GPU Computational Accelerator for AI Machine...	$879.00	Buy on Amazon
2		PRIVATE AI INFRASTRUCTURE ON LINUX: Build, Secure, and Operate Self-Hosted LLM Servers with Ollama,...	$8.99	Buy on Amazon

As an Amazon Associate I earn from qualifying purchases.

What Baseten Dedicated Inference provides

Dedicated Inference is designed for organizations that need a private serving environment for a model they are licensed to run. Baseten describes a split architecture: its control plane handles orchestration and infrastructure management, while each customer model runs in an isolated workload plane with dedicated GPU resources. Scheduling and compute are not shared with another customer’s model deployment.

The isolation changes the procurement and engineering path. You are not selecting Gen-4.5 or Aleph 2.0 from a generic public catalog and receiving a common multi-tenant API key. Instead, the deployment is created for your organization after the model-weight license is in place. Baseten can reserve GPU capacity for the workload and autoscale it when demand rises, while your application calls the resulting private service.

#1 Best Overall

Sale

HPE NVIDIA Tesla V100 32GB HBM2 PCIe 3.0 x16 Passive GPU Computational Accelerator for AI Machine Learning HPC Deep Learning 699-2G500-0216-400 (Renewed)

• NVIDIA Volta GV100 Architecture — 4,608 CUDA Cores, 640 1st-Gen Tensor Cores delivering 14 TFLOPS FP32 and 112 TFLOPS deep learning performance for AI training, inference, HPC, and scientific computing workloads
• 32GB HBM2 ECC Memory — 900 GB/s Bandwidth — High-bandwidth memory on a 4096-bit bus with ECC error correction provides the memory capacity and throughput required for the largest AI models, simulations, and datasets
• PCIe 3.0 x16 Interface — 250W TDP — Standard PCIe Gen3 connectivity with passive cooling designed for enterprise rack server deployment in HPE ProLiant, Dell PowerEdge, and Supermicro platforms with adequate chassis airflow
• NVLink — Scale to 96GB Unified Memory — Connect two V100 GPUs via NVLink at 300 GB/s bi-directional bandwidth to scale GPU memory from 32GB to 96GB for larger AI training and HPC workloads
• Multi-Precision Computing — Supports FP64 (7 TFLOPS), FP32 (14 TFLOPS), FP16 (112 TFLOPS) and INT8 precision modes for flexible deployment across training, inference, and scientific simulation workloads

$879.00

Buy on Amazon

What this arrangement is useful for

• Keeping model execution in a single-tenant environment for internal or customer-facing production workloads.
• Using dedicated GPU capacity and autoscaling instead of maintaining your own inference cluster.
• Applying organization-specific data handling, access controls and deployment practices.
• Fine-tuning on your own data when the direct Runway license allows it.
• Obtaining support across customization, deployment and production serving from Runway and Baseten.

Baseten states that its dedicated deployments are covered by its SOC 2 Type II and HIPAA programs. Those programs describe the provider’s controls; your team still has to configure identity, retention, logging and application-level safeguards correctly.

Choose the right Runway model

Gen-4.5 and Aleph 2.0 solve different video problems. Gen-4.5 creates a new shot from a text or image starting point. Aleph 2.0 takes existing visual material and performs a targeted in-context edit. Selecting the model by the desired operation, rather than by the model’s name, avoids unnecessary cost and quality loss.

The per-second figures are Runway Dev direct-API prices. They are not a statement that a Baseten deployment includes infrastructure or that Baseten bills by generated-video second.

When Gen-4.5 is the better fit

Use Gen-4.5 when the source material is an idea, a still image or a brief visual reference and you need a fresh moving shot. A product team can animate a pack shot, a creative team can test alternate campaign scenes, and a previsualization team can turn a written concept into a short clip. The catalog’s 720p output and ten-second maximum should be part of your delivery plan: longer sequences need to be designed as multiple shots and assembled downstream.

When Aleph 2.0 is the better fit

Choose Aleph 2.0 when you already have footage and want to preserve most of its composition while changing a defined attribute. Examples include replacing a location, changing clothing color, adding a weather condition, relighting a scene or applying a period-film look. Its catalog allows clips up to 30 seconds and says the output matches the input, which makes it more suitable than a fresh-generation model when continuity with the original material matters.

Requirements before you request a deployment

Secure the model-weight license

Baseten’s announced process requires customers to license the Runway model weights directly from Runway. The license determines whether you may use the weights commercially, fine-tune them, and serve the resulting deployment to your users. Do not begin production provisioning on the assumption that a Runway Dev subscription automatically grants the rights needed for a private Baseten workload; obtain the written commercial scope from Runway.

Define the workload

Prepare a short specification for both the model and the application that will call it. Include expected requests per minute, peak bursts, clip duration, input resolution, output resolution, acceptable queue time, regions in which data may be processed, retention requirements and whether you need fine-tuning. These details affect GPU selection, autoscaling thresholds and the security review.

Decide who owns each integration layer

Assign an owner for the Runway licensing relationship, a Baseten platform owner and an application owner. The platform owner should control deployment credentials and secrets; the application owner should handle asynchronous jobs, retries, result storage and user-facing status. Keeping those roles separate prevents a model key from being embedded in a client application.

Deployment procedure

1. Confirm rights and commercial scope with Runway

1. Describe the intended product, geography, users and expected volume to Runway.
2. Confirm that the agreement covers deployment of the Gen-4.5 or Aleph 2.0 weights on a third-party dedicated inference platform.
3. Ask specifically whether fine-tuning on your data is permitted and what restrictions apply to the resulting model.
4. Record the approved model version, permitted inputs and outputs, retention obligations and any attribution or redistribution requirements.

Baseten cannot replace this licensing step: the provider supplies infrastructure, while Runway controls the rights to the weights and their permitted uses.

2. Engage Baseten Dedicated Inference

1. Request a single-tenant workload plane and explain the GPU, region, networking and compliance requirements from your workload specification.
2. Choose reserved capacity appropriate to normal traffic and define the burst behavior that autoscaling should handle.
3. Establish the control-plane administrators, deployment environments, logging policy and secret-management method.
4. Have Baseten confirm the service’s endpoint format, authentication scheme, rate limits, health checks and rollout procedure in writing.

Those endpoint and authentication details are onboarding outputs. The October 2026 announcement does not provide a universal URL or header name, so a code sample that invents them would not be reliable.

3. Deploy and customize the licensed model

1. Provide Baseten with the Runway licensing information and the model artifact or transfer instructions requested by the onboarding team.
2. Choose whether to deploy the base model or an approved fine-tuned variant.
3. Configure GPU type, minimum warm capacity, autoscaling limits, timeouts, concurrency and storage according to your workload.
4. Run a private acceptance test covering representative prompts, input sizes, failure responses, output retrieval and cancellation behavior.
5. Promote the accepted revision through staging and production using the deployment and rollback process supplied by Baseten.

Runway and Baseten describe support from customization through production serving. Treat a fine-tuned model as a separate release: keep its training-data provenance, evaluation results and license scope alongside the deployment record.

4. Connect your application

Runway’s public SDK documentation demonstrates the request concepts for Gen-4.5: identify the model, provide a prompt, choose an aspect ratio and duration, submit an asynchronous task, wait for completion and then read the returned video URL. This illustrates the application flow, not a Baseten endpoint. Once onboarding supplies the private service contract, map these concepts to the issued URL, headers and deployment name.

A Runway Dev JavaScript example uses the following fields:

const task = await client.imageToVideo.create({
  model: 'gen4.5',
  promptText: 'A slow cinematic push-in on a glass sculpture at sunrise',
  ratio: '1280:720',
  duration: 5
});

const output = await client.tasks.waitForTaskOutput(task.id);
const videoUrl = output.output[0];

The corresponding Python flow uses the same model, promptText, ratio and duration values and waits for task output before reading the video URL. In a Baseten integration, keep the asynchronous pattern but use the private endpoint and authentication mechanism that Baseten documents for your deployment. Do not copy the Runway Dev hostname or assume that its credentials authorize a Baseten service.

5. Put production operations and governance in place

1. Store service credentials in a server-side secret manager and issue the application only the minimum permission it needs.
2. Track task IDs, submission time, model revision, prompt, input reference, completion status and output location without placing sensitive media in ordinary logs.
3. Set client and server timeouts that allow for queueing and generation, and implement bounded retries for transient failures rather than resubmitting every timeout blindly.
4. Measure queue latency, generation latency, GPU utilization, error rate, cancellation rate and output-storage growth.
5. Define a rollback target and test it before a model revision or fine-tune becomes the only production option.
6. Document deletion, retention and export procedures for uploaded video, generated output and task metadata.

Gen-4.5 request design

Gen-4.5’s public quickstart is an image-to-video flow, so your application should treat the still image and prompt as versioned inputs. Keep the ratio and duration explicit instead of relying on defaults. A 1280:720 ratio in the example requests a 16:9 output; your Baseten deployment may expose the same choices, but the accepted values must be confirmed during onboarding.

Build a reliable asynchronous job loop

• Create a server-side job record before submitting the generation request.
• Persist the provider task identifier immediately after submission.
• Poll or use the callback mechanism documented for your deployment, with exponential backoff and a maximum wait window.
• When complete, copy the returned video into storage you control and save its checksum and expiration information if the provider URL is temporary.
• Mark permanent failures separately from user cancellations so operators can distinguish capacity problems from invalid input.

The public example’s waitForTaskOutput() call is convenient for a script, but a production web service should not hold a request thread open for the entire generation. Return a job status to the client and let a worker process completion.

Rank #2

PRIVATE AI INFRASTRUCTURE ON LINUX: Build, Secure, and Operate Self-Hosted LLM Servers with Ollama, Open WebUI, vLLM, Docker, RAG, and GPU Acceleration

$8.99

Buy on Amazon

Aleph 2.0 prompting and editing

Runway’s guidance is that “There are no strict rules for prompting with Aleph 2.0, but simplicity, iteration, and precision work best.” The practical implication is to name one action and one target before adding optional style information.

Use an explicit action verb

Start with a verb such as add, remove, change, replace, re-light or re-style. Follow it with the exact element and the desired result. Official examples include:

• change background to ancient city ruins
• Change outfit to bright yellow fur coat
• heavy overcast and raining
• Chiaroscuro lighting
• 1920s silent era film

For a controlled edit, identify what must remain unchanged in your own prompt, such as the subject’s position, camera movement or facial expression. Avoid combining unrelated transformations in the first pass; make one edit, inspect the result and then iterate.

Preserve continuity deliberately

Because Aleph 2.0 edits existing material, review the first and last frames for identity, geometry, lighting and motion continuity. If a background replacement also changes the actor’s clothing, simplify the prompt and submit separate edits. Keep the original clip and every generated revision so that an unwanted change can be reversed without regenerating the source.

Pricing and capacity planning

There are two different cost layers. Runway lists direct-API model prices, while Baseten bills dedicated infrastructure by compute minute. A Baseten deployment does not mean the infrastructure is included in the Runway Dev per-second figures.

Baseten’s pricing page lists Basic, Pro and Enterprise plans for dedicated deployments. Ask for a written estimate that separates plan charges, GPU time, storage, networking, support and any minimum reservation. Compare the estimate with your expected warm capacity and burst profile, not just the number of completed clips.

Capacity choices that affect the bill

• Warm minimum: keeping GPUs ready reduces cold-start delay but incurs compute cost during idle periods.
• Autoscaling ceiling: a high ceiling absorbs bursts but can multiply spend if a client retries aggressively.
• Concurrency: allowing too many simultaneous jobs can increase queueing or exhaust memory; benchmark the deployment with your actual clip sizes.
• Retention: generated videos and source uploads can become a larger storage and egress cost than expected if deletion rules are absent.

Security, compliance and data handling

Baseten describes isolated compute and scheduling for each dedicated deployment and supports deployments under SOC 2 Type II and HIPAA programs. Use those assurances as a foundation for your review, then verify the exact controls and regional availability in your contract and onboarding materials.

Questions for the security review

• Where are inputs, intermediate files, logs and outputs stored, and how long are they retained?
• Can administrative access be restricted by role, network or single sign-on?
• Which events are auditable, and can logs be exported to your monitoring system?
• What encryption is used in transit and at rest, and who manages the keys?
• How are model updates, fine-tuned artifacts and rollback images protected?
• What happens to queued jobs and stored media when a deployment is deleted?

For regulated footage or personal data, minimize prompt and media content in logs, redact user identifiers where possible and enforce deletion at both the application and provider layers.

Troubleshooting common deployment problems

“We have a Runway key, but Baseten rejects the request”

A Runway Dev credential is not proof of authorization for a Baseten-hosted model. Verify that the model-weight license is complete, that Baseten has the deployment mapped to your account and that your application is using the private endpoint and header format issued during onboarding.

“The endpoint works, but jobs remain queued”

Check the deployment’s minimum GPU count, autoscaling ceiling, current concurrency and regional capacity. A queue can reflect an intentionally low warm minimum rather than a model failure. Compare queue latency with GPU utilization before increasing retries.

“Aleph changed more of the scene than requested”

Reduce the prompt to one action and one target, state the surrounding details that must remain, and iterate from the original clip. Combining a wardrobe change, a weather change and a style change in one request makes it harder to control which visual attributes move.

“Generated output is lost after completion”

Do not treat a returned provider URL as permanent storage. Download or copy the asset to your controlled object store as part of the completion worker, save the task ID and checksum, and apply your retention policy there.

“Costs spike during a campaign”

Inspect autoscaling events, duplicate submissions and client retry behavior. Add idempotency at your application layer, cap concurrent jobs per tenant, and set an autoscaling ceiling before raising the warm minimum.

How to decide between the two models on Baseten

What you should request in the onboarding packet

Before writing production code, obtain a complete service contract from Baseten and Runway. It should state the approved model and revision, endpoint URL, authentication headers or token exchange, request and response schemas, accepted media formats and size limits, asynchronous status method, timeout and retry guidance, rate limits, error codes, callback or polling behavior, regional placement, retention policy, GPU configuration, support contacts and the rollback process.

That packet closes the gap between Runway’s public SDK concepts and your private deployment. Until it arrives, use the public Gen-4.5 example only to design your application’s job model and UI states; do not hard-code an endpoint, header or deployment command that the published announcement does not specify.

Bottom line

Baseten makes Gen-4.5 and Aleph 2.0 available as privately deployed, single-tenant workloads for customers that obtain the appropriate Runway license. Gen-4.5 is the choice for creating new shots; Aleph 2.0 is the choice for precise edits to existing video. The practical implementation is license first, provision dedicated GPUs through Baseten, deploy and test the model, then integrate using the endpoint and authentication details supplied during enterprise onboarding. Budget separately for Runway’s model usage and Baseten’s compute-based infrastructure.

Quick Recap

SaleBestseller No. 1

HPE NVIDIA Tesla V100 32GB HBM2 PCIe 3.0 x16 Passive GPU Computational Accelerator for AI Machine Learning HPC Deep Learning 699-2G500-0216-400 (Renewed)

$879.00

Bestseller No. 2

PRIVATE AI INFRASTRUCTURE ON LINUX: Build, Secure, and Operate Self-Hosted LLM Servers with Ollama, Open WebUI, vLLM, Docker, RAG, and GPU Acceleration

$8.99

Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.

AI video generationAleph 2.0BasetenDedicated Inferenceenterprise AIRunway Gen-4.5video editing API