Muna Shared · Serverless inference for your own models

Bring your model.
Pay only while it serves requests.

Your models on shared B200s from $3/hour. No GPU bill while idle, cold starts under 10 seconds, and an OpenAI-compatible API.

Join the waitlist See pricing

$3/hr
B200, while generating: $0
while idle: <10s
cold starts, not minutes: OpenAI
compatible API

The problem

You're paying for GPUs that are doing nothing.

Warm containers bill all day

Busy 2 hours, billed for 24. Most of your GPU bill is waiting for traffic.

Scale to zero is a trap

Minute-long cold starts mean real users wait. So you keep a GPU warm anyway.

Every model wants its own GPU

One dedicated GPU per model, whether it's busy or not. The bill adds up fast.

How it works

Change two lines. Compile. Call it like OpenAI.

01Write a Python function

Load your model, return its outputs. Starting from one of our functions? Change two lines.

# Load a model with 🤗 transformers
model = AutoModel.from_pretrained("Qwen/Qwen3.8-27B")

@compile(
    tag="@your-org/your-model",
    ...
)
def qwen_3_8_27b(messages: list[Message], ...) -> Iterator[ChatCompletionChunk]:
    ...

02Compile it

One command turns it into a native binary, tens of megabytes. No container, no Python at runtime.

$ muna compile qwen_3_8_27b.py

03Call it like OpenAI

Scales to zero. Wakes in under 10 seconds. Billed only while it generates.

from openai import OpenAI

client = OpenAI(base_url="https://inference.muna.ai/v1", api_key="...")
stream = client.chat.completions.create(
    model="@your-org/your-model",
    messages=[{ "role": "user", "content": "Hello!" }],
    stream=True
)

04Share the GPU, split the bill

Your model shares a GPU with others and only holds it while it has work.

Dedicated: one GPU per model

2 GPUs at $14/hr, 70% of each one unused

30% utilized

Co-located: models take turns

1 GPU at $7/hr, same traffic

60% utilized

Model A

Model B

Idle

Pricing

Pay for work, not for waiting.

Per-second billing while your model has requests running. Overlapping requests bill once.

GPU	Memory	Muna Shared	Modal	Rate difference
B300	288 GB	$3.75/hr	$7.10/hr	−47%
B200	180 GB	$3.00/hr	$6.25/hr	−52%
H200	141 GB	$2.75/hr	$4.54/hr	−39%
H100	80 GB	$2.25/hr	$3.95/hr	−43%
RTX PRO 6000	96 GB	$1.95/hr	$3.03/hr	−36%

Modal list prices from modal.com/pricing. The lower rate is the smaller part of the savings: under heavy sharing your model runs up to 2× slower, which can bring the cost of the same work close to Modal's hourly rate. The savings that hold either way come from never paying for idle time, model loading, or a keep-warm window.

The fine print

Shared means shared.

Any model, any traffic

Spiky or steady, small or large. Either way, you pay for the seconds your model runs.

Fair share of the GPU

When neighbors are busy, throughput is shared among co-tenants, down to about 1/2 of a GPU to yourself. Past that, we move models apart.

	Billed?
Prefill and generation	Yes
Time slowed by other models on the GPU	Yes
Waiting in the queue before your request starts	No
Loading your model (cold starts)	No
Idle time and keep-warm windows	No
CPU, memory, and storage	No

Calculator

See what you'd pay.

GPUHours serving requests per dayCold starts per dayModal cold start (seconds)Modal idle window (seconds)

Modal, kept warm: $4,500
Modal, scale to zero: $1,141
Muna Shared: $360–$720

At least 37% below Modal's scale-to-zero bill, even under the heaviest sharing. Per month. The Muna range runs from a GPU to yourself to the heaviest sharing (2× slower).

Modal scale to zero bills each cold start and the idle window after each burst (Modal's default is 60 seconds). It also means 50 requests a day wait about 90 seconds for a model to load. Muna bills neither, and cold starts take under 10 seconds.

Live from production

Cold starts under 10 seconds.

Model	Load-to-ready @p50 (@p90)
`@google/gemma-4-26b-a4b-it`	8.2 s (11.6 s)
`@orcarouter/qwen-3.8-27b-uncensored`	21.8 s (37.0 s)
`@qwen/qwen-3.8-27b`	9.1 s (13.2 s)
`@nomic/nomic-embed-text-v1.5`	1.7 s (2.5 s)
`@qwen/qwen-3-embedding-8b`	2.7 s (4.1 s)

Load-to-ready is the time from a GPU starting to load the model to the model being ready to serve, measured in production and refreshed every five minutes. The first token of a cold request also waits for prefill, which depends on prompt length. Larger models take longer to load.

FAQ

Before you sign up.

Is this a GPU I can SSH into?

No. Muna Shared is serverless inference for any model you bring. You get an API endpoint, not a machine.

Which models are supported?

Any model you can write as a Python function: LLMs, embedding models, and more. Ready-made functions cover Qwen 3.8 27B, Qwen 3 4B, Gemma 4 26B A4B, Gemma 4 31B, GPT-OSS 20B, Phi 4 Mini. Stuck? Ask.

Does anything cost extra?

No. You pay the hourly rate while your model is generating, and nothing else.

What if I need guaranteed speed?

Dedicated GPUs run your model alone, with no sharing slowdown. Talk to us.

Early access

Get in early. Lock in the price.

The first 20 teams keep launch pricing for 12 months, with an engineer on Slack for their first compile.

Bring your model.Pay only while it serves requests.