Run your real workflow on Minima-optimised models with up to $1,000 in free credits. Measure tokens per second, latency and cost per task against what you run today. Scale only when the numbers hold.
Weight and KV-cache optimisation for your own models
Custom serving runtime tuned to your GPUs
2 to 4 week benchmark with quality, throughput, memory and cost gates
Private model and data path, deployment support
Prices in US dollars.
For neoclouds & inference operators
Get more revenue-producing inference from every GPU you already own.
Minima integrates at the model and runtime layer of your serving stack, so the same fleet serves more tokens, holds larger models and competes on price per token.
· Benchmark on a representative model· Quantified extra capacity and margin· Offer Minima-optimised models to your customers
All prices are in USD per 1 million tokens. Speeds are indicative and may vary by workload and configuration.
Free Test credits are spent at these same rates, so your trial bill is your production bill.
What you actually pay for
Performance first. Economics second.
You don't buy compression ratios. You buy a model that is fast enough to ship, and a GPU bill that doesn't grow with every new user.
Feature
Benefit
Value to you
Custom runtime and speculative decoding
Up to 2× more tokens per second on the same model
Ship the open model you prefer instead of falling back to a closed API.
3.5× KV-cache compression
Longer context and higher concurrency per GPU
Hold your latency SLA as traffic grows.
4× smaller model weights
The same model fits on fewer nodes
Lower cost per accepted task, at the quality bar you set.
Minima internal benchmarks. Configuration-specific results, not a general performance guarantee. Minima internal benchmark, Qwen3.8-2.4T on NVIDIA B200. Results depend on the model, workload and serving configuration.
How credits work
Commit. Consume. Top up. Upgrade.
01
Commit
Pick a monthly credit pool, or stay on pay as you go. Commits get 15% off every rate.
02
Consume
Every request draws from your pool at the per-model rate. You see TPS, latency and spend per workload.
03
Top up
Running low? Add credits at the standard rate in one click. No renegotiation, no paused traffic.
04
Upgrade
Topping up every month? Move to a larger commit and a lower effective rate.
For enterprise
Benchmark your model before you sign.
Bring one real workload. We show you the before and after on your hardware, against the quality bar you set. If it doesn't hold, you know in weeks, not quarters, and you keep the readout.
2 to 4
weeks from workload to readout
1
real workload, measured end to end
Week 0
You share the workload
Model, GPUs, context, concurrency, traffic shape, quality and latency constraints.
You get a scoped plan with explicit gates.
Weeks 1 to 3
We optimise and serve it
Weights, KV cache and runtime tuned for your model on your hardware.
You get your model running on mnma.
Readout
Your team runs it
Real traffic, side by side with your current baseline.
You get TPS, latency, peak VRAM and cost per task, before vs after.
Your call
You decide
Move to private production on an annual licence, or keep the insight.
You get what production takes and what it costs.
Self-serve test
Run your workflow on hosted models with free credits.
$0up to $1,000 in credits
Managed private benchmark
Your model, your GPUs, fixed scope, production-shaped result.
$10,000[credited to year-one licence]
Production rollout
Private deployment, then more models and workloads.
Credits are in dollars. $1,000 in credits buys $1,000 of inference at the published per-model rates, so what you see in the test is what you pay in production.
Why "up to" $1,000?
You get starter credits the moment you sign up. The rest unlocks after a short check of the workload you want to test, so the credits go to real production use cases.
Do I need to change my code?
No. The API is OpenAI-compatible. Point your client at the Minima endpoint and swap the key.
What happens when free credits run out?
Requests pause until you add a payment method or a commit. Your keys, settings and usage history stay.
Can we run Minima in our own cloud?
Yes. The Enterprise plan deploys the mnma Engine on-premises or in your VPC, with a private model and data path.
How do you prove the gains?
Every result is tied to a model, hardware, workload and quality gate. Change the setup and the result can change, which is why we benchmark on yours.
Do you store or train on our data?
[Data retention and training policy.]
Tell us what you run today
See what Minima can do for your model and GPUs.
Start on free credits today, or send us a workload and we'll propose a benchmark with explicit quality, throughput, memory and cost gates.
