Custom logic on pre-hosted models

If your goal is to run custom application logic directly alongside pre-hosted open-source foundation models, you should look at inference platforms that go beyond a standard model API. The key distinction is whether a platform lets you execute your own code in the inference service, or simply lets your application call a hosted model.

Here are the strongest candidates based on their documented capabilities.

Platforms to evaluate

• 

1. Hugging Face Inference Endpoints

(Custom inference logic)

Supports custom Python handlers for preprocessing, postprocessing, additional request parameters, and other application-specific logic. You can start from an existing model on the Hugging Face Hub and deploy it to managed, autoscaling infrastructure.huggingface.co+1

Best for: Python-based applications where you want custom logic inside the model-serving process.

Docs:Custom Inference Handler

---

• 

2. Replicate

Lets you package arbitrary code and model weights into a deployable service using Cog, then expose it through an API. Managed deployments handle GPU infrastructure, scaling, and updates.replicate.com+1

Best for: Shipping a self-contained inference application quickly, especially when you need more than a standard LLM endpoint.

Docs:Deploy a custom model

---

• 

3. Together AI

Offers serverless inference for open-weight models, plus dedicated model and container inference options. Its dedicated container offering is particularly relevant for custom inference pipelines and non-standard runtimes.www.together.ai+1

Best for: Getting a model running quickly, then moving to a more customized managed serving environment if needed.

Important distinction: Verify whether your intended application logic can run inside the container or must live in a separate service.

Docs:Serverless Inference

---

• 

4. Fireworks AI

Provides managed, optimized inference for open models and customized model variants, with serverless and dedicated deployment options. Its main strength is high-performance model serving.docs.fireworks.ai+1

Best for: Low-latency, high-throughput LLM applications.

Important distinction: Confirm support for arbitrary application code within the serving endpoint; its standard inference API alone does not establish that capability.

Docs:Fireworks documentation

• 

5. Modal

A particularly strong fit for your stated requirement. Modal lets you deploy Python code, load open-weight models, run preprocessing and postprocessing, execute agent tools, and serve a custom inference stack on managed GPUs. It also supports autoscaling and scale-to-zero workloads.modal.com+1

Best for: Combining model inference and application logic in one deployable Python service without managing GPU infrastructure.

Docs:Modal Inference

---

• 

6. Baseten

Its Truss framework packages custom Python model code, dependencies, and weights into a managed deployment. Baseten also offers optimized inference engines and Chains for composing multi-step workflows, including model calls and postprocessing.github.com+2

Best for: Production applications that need custom model behavior, optimized serving, and potentially multi-model workflows.

Docs:Baseten model deployment

Which should you choose?

What to check before committing

The phrase “custom application logic” can mean different things. Make sure the platform supports the specific execution model you need:

• Inline execution: Your Python or application code runs in the same service as the model, rather than making a second network call to another service.
• Preloaded model weights: The foundation model is loaded once and reused across requests instead of reloaded on every invocation.
• Fast deployment: Existing model weights, dependencies, and serving infrastructure are reusable, minimizing setup and deployment time.
• Scaling and latency: The service can handle concurrent requests without excessive cold starts or duplicated model loading.
• Custom dependencies and security: You can install your libraries and configure secrets, network access, and isolation appropriately.

My recommendation: Start by comparing Modal, Hugging Face Inference Endpoints, and Baseten. They provide clear paths to custom inference behavior, while Modal is especially attractive if you want to write the whole service in Python. If you mainly want to call a pre-hosted model through an API, rather than execute code inside its serving environment, Together AI or Fireworks AI may be simpler.

One question that would narrow this down considerably: Do you want to deploy arbitrary Python code in the same runtime as the model, or do you want a hosted model API with custom logic executed as part of a managed workflow?