Your models. Your infrastructure as code. From zero to thousands of replicas, Infimal handles demand you can’t predict. Pay for the work that runs.
infimal. Set up your key →curl -fsSL https://infimal.ai/install.sh | sh client.model(id) returns a handle. Chain the options that surface takes, then run() or stream() it. No infrastructure to declare first.
An id from GET /v1/models. No traffic forecast, no capacity to reserve before the first call.
Chat takes system, temperature, max_tokens. Images and video take seed, steps, guidance, LoRAs, an uploaded reference.
run() waits for you. submit() returns a job your process can lose and pick back up with client.job(id).
Images and video are durable jobs, not one blocking call. Submit, disconnect, and adopt the same job by id whenever you come back for it.
Read the options by surfaceclient.model(id) returns a handle; .chat() builds one request. run() streams the completion under the hood and hands back the assembled text and usage.
from infimal import Client client = Client.from_env()chat = client.model("qwen3-8b") result = ( chat.chat("Explain this diff.") .system("Answer in one paragraph.") .temperature(0.4) .max_tokens(256) .run())print(result.text, result.usage, result.request_id)max_tokens caps output tokens, not a dollar amount. system() replaces a leading system message, or inserts one.
You don’t know how busy your model will be. Compare the full monthly serving bill for one workload, then change the usage.
Hold enough capacity for the incoming stream, through busy days and quiet ones.
1 × H100 SXM × 730 hr × $4.29/hr
of engineering time for serving, scaling, access, and monitoring.
Process the same audio. Retain the initialized model between bursts of work.
657,000 audio seconds × $0.0010/sec
Declare your infrastructure. Deploy your ready-to-run model.
Both sides process the same checkpoint, audio duration, and peak of one real-time stream. The daily trace is an illustrative distribution of that exact audio volume; the last day contains 10 hours to make a 730-hour month. The rental is sized with the throughput you enter, rounded up to whole GPUs, and held for 730 hours. Infimal is priced per active audio second. Equal output quality and real-time service are assumptions to validate for your model.
The default $4.29/GPU-hour is Lambda’s published single-H100 SXM rate, checked September 18, 2026. The $0.0010/second this starts at is a placeholder you replace, not a catalogue price: the catalogue prices models per token or per request, so there is no per-audio-second rate on it to quote. The one-stream/GPU throughput is an input too, not a benchmark. Both estimates exclude tax, storage, egress, build charges, and engineering labor. An autoscaled rental is a different operating strategy.
Why start at 25%? This is a planning estimate for an endpoint with variable demand: 182.5 active audio hours in a 730-hour month. Published evidence supports underutilization but does not establish an average monthly usage for model companies. Anyscale reports production GPU utilization often below 50%; Cast AI reports 5% in its enterprise sample, which excludes AI labs. Those are GPU-compute measures, not this model’s active-audio fraction. The 25% default is our assumption; use your traffic to refine it.
Setup times are planning estimates for the same working model with known dependencies, from infrastructure configuration to a first serving endpoint with scaling, access, and monitoring. They exclude model development, data preparation, and artifact transfer time. Existing infrastructure and model requirements can change either estimate. Setup time is shown separately and is not included in the monthly dollar totals.
Serving costs and setup time, side by side. Adjust the assumptions for your model.
Restore initialized state when work returns, instead of keeping every model on a rented GPU.
Infimal places and scales replicas within your declared bounds. Your team keeps shipping.
The same infrastructure definition covers the experiment, the launch, and the quiet hours after.
Tokens, streaming seconds, and request priority. Your model determines the rate.
Create in Studio, or build with the SDK and CLI.