Concurrent Unit Limits
How plan concurrent units translate into parallel inference capacity.
For fixed monthly pricing plans, Featherless uses reservations of capacity, that determine your model size limit and how many requests can run in parallel.
Individual fixed plans include a fixed concurrent unit budget for interactive use. Business plans make that budget explicit and scalable, from a few concurrent units to as many as your workload needs.
For plan-specific model limits, included concurrent units, and availability, see the plans.
Model Concurrent Unit Costs
Each model has a concurrent unit cost based on its size.
When a request starts, it reserves that model's concurrent unit cost from your plan budget. The units are released when the request completes or is cancelled, so higher-cost models reduce the number of requests that can run in parallel.
How concurrent units are reserved
This example uses a 4-unit plan budget. Units are reserved while requests are in flight and released when they finish.
Example plan budget
4 concurrent units
Small model request
1 unit per request.
Larger model request
2 units per request.
Very large model request
4 units per request.
More concurrent units mean more parallel requests. Exact unit cost is set per model class and shown on model pages.
Model category | Concurrent Unit Cost | Examples |
|---|---|---|
Small models (less than 16B) | 1 | Qwen, Llama |
Larger models (less than 34B) | 2 | Qwen Coder, Mistral |
Very large models (70B or more) | 4 | MiniMax, Kimi, DeepSeek |
There are some exceptions. Exact concurrent unit cost is set per model class and shown on model pages.
How to read your plan
Your plan shows a concurrent unit budget. Each request reserves units based on the model category while it is running.
For example, with a 4-unit budget you can run:
4 small-model requests at the same time
2 larger-model requests at the same time
1 very-large-model request at the same time
Any mix that adds up to 4 concurrent units
With an 8-unit budget, double those examples.
For scalable plans, multiply the included concurrent units by the number of plan units you have purchased.
Requests that would take your total reserved units above your plan budget are rejected with HTTP 429.
Last edited: Jul 16, 2025