TOKEN FACTORY

Open models. Production ready.

Enterprise-scale inference for open models, from open weights to production delivery. Build with an OpenAI-compatible API on Aolani’s managed GPU infrastructure.

Overview

☷   GET STARTED

Build on your API gateway in minutes

A focused home for keys, balance, routing, and service health.

Ready for your first request
First API requestsa-agentic-coding
● ● ●curl https://<gateway-host>/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer sk-demo••••" \
+  -d '{"model":"nvidia/nemotron-3-ultra-550b-a55b","messages":[…]}'

Interactive Aolani console overview: API onboarding and platform shortcuts.

PLATFORM

From open weights to production delivery.

A single managed platform for model access, endpoint operations and production assurance.

NVIDIA ECOSYSTEM

H100 / GB200 / GB300-class + Dynamo

72 GPUs · One rack in service

Built on NVIDIA Dynamo, the open-source distributed inference framework for production GPU inference across H100, GB200, and GB300-class systems. The technology path follows NVIDIA’s roadmap.

Expand from H100 through GB200 and GB300-class systems to next-generation platforms. Hardware upgrades stay transparent at the API layer, so applications do not need to change.

Reported platform validation maintained high-concurrency throughput during elastic rescaling. Capacity is designed to grow linearly as resources are added.

01

Inference service

Text, code and reasoning models through an OpenAI-compatible API. Choose shared or dedicated endpoints.

02

Endpoint and model operations

Day-0-grade onboarding for new models, with a serving stack optimized per model and dual inference engines.

03

Operations assurance

Full lifecycle management from planning to deployment and operations, monitoring and alerting, a 99.9% SLA and repeatable delivery processes.

CORE CAPABILITIES

Build the product. We’ll run the inference.

A complete operating foundation for production AI, with model serving, scaling, security and support in one place.

Open models in production

Serve mainstream open models through an OpenAI-compatible API. Scale compute with demand, with prefill and decode scaling independently.

Scale rack by rack

Start with the configuration your workload needs—from H100 through GB200 and GB300-class infrastructure. Add capacity as demand grows, with a path to next-generation GPU platforms.

Clear token pricing

Pay per million tokens, with separate input and output rates. No hidden GPU or infrastructure fees; dedicated endpoints are priced by capacity.

Long context and agents

Work with million-token-scale context, native reasoning and function calling. Keep reasoning separate from the main response for a clear output structure.

Bring your own models

Deploy custom checkpoints directly to endpoints with the same security and performance guarantees as standard endpoints. Aolani manages the infrastructure behind your model.

Fully managed operations

Keep production covered with monitoring, alerting and closed-loop asset and incident management. A 99.9% SLA supports operations while Aolani manages the GPU infrastructure.

MODEL CATALOG

The open models your product depends on.

Six current open-weight model families for production workloads, with additional models available on request.

Explore model selection and request parameters in the Aolani playground. Responses are authored examples; no model is called.

1M context · reasoning

GLM 5.2

Large-scale reasoning for long-horizon agents, project-level software engineering, and complex multi-step automation.

1M context · multimodal · 2.8T

Kimi K3

A 2.8T open-weight multimodal reasoning model for complex coding, knowledge work, and long-horizon agent workflows.

1M context · 49B active MoE

DeepSeek V4 Pro 0423

A 1.6T-parameter sparse MoE with 49B active parameters for advanced reasoning, coding, and long-horizon agent workflows.

1M context · 95B active MoE

Qwen3.8 2.4T A95B

An open-weight 2.4T sparse MoE with 95B active parameters for coding, research, complex reasoning, and agentic workflows.

262K context · 55B active MoE

NVIDIA Nemotron 3 Ultra

NVIDIA’s 550B hybrid Transformer–Mamba MoE, with 55B active parameters for orchestration, coding agents, deep research, and enterprise tasks.

131K context · 5.1B active MoE

OpenAI gpt-oss-120b

OpenAI’s 117B open-weight MoE, activating 5.1B parameters per pass for reasoning, agents, tool use, and structured outputs.

ENDPOINT OPTIONS

Same model. Same outputs. Your performance profile.

Match the endpoint configuration to concurrency, responsiveness and cost. The figures below are reported configuration results; validate them against your model and traffic.

8 GPUs

Standard

1 Prefill + 1 Decode

4,600+ tok/s at high concurrency · 94 tok/s single-stream

For concurrent workloads, batch processing and cost efficiency.

16 GPUs

Performance

2 Prefill + 2 Decode

6,200+ tok/s at high concurrency · 120 tok/s single-stream

For low-latency chat and retrieval-augmented generation with long documents.

Need dedicated capacity?

Whole-node isolation, reserved capacity, bring-your-own-model support and a 99.9% SLA. Configure resources and scaling policies around your traffic profile.

ENGINEERED FOR INFERENCE

Performance comes from the whole stack.

Optimizations are applied per model, and enabled only after quality preservation is confirmed.

6,200+

Tokens per second

Reported aggregate throughput · 16 GPUs

6,200+

Tokens per second

Reported aggregate throughput · 16 GPUs

9 sec

150K-token document

Reported long-context parsing result

9 sec

150K-token document

Reported long-context parsing result

99.9%

Service-level agreement

Monitoring and incident management

99.9%

Service-level agreement

Monitoring and incident management

01

Prefill-decode disaggregation

Reported validation preserved high-concurrency throughput while prefill and decode scaled independently.

02

KV and prefix caching

Reported cache hit rates of 99–100% on repeated prefixes.

03

NVFP4 quantization

4-bit weights and activations make room for larger models and longer contexts within the same memory.

04

Continuous batching

Reported aggregate throughput of 6,200+ tok/s on 16 GPUs.

05

Cross-node communication fusion

Up to approximately 1.9× faster single-stream decoding in multi-node tensor-parallel scenarios.

06

Million-token context

A 150K-token document parsed in 9 seconds in the reported validation.

Reported validation also reached 100% request success and a peak of 7,600+ tok/s. The peak is separate from the documented 16-GPU configuration result. Model, workload and test conditions affect performance; review your requirements with Aolani.

DEVELOPER EXPERIENCE

Your existing API workflow. A new inference foundation.

Use /v1/chat/completions with streaming and the full vLLM parameter set. Keep the familiar OpenAI client and change the endpoint.

PYTHON / OPENAI SDK

from openai import OpenAI

client = OpenAI(
    base_url="https://<gateway-host>/v1",
    api_key="YOUR_API_KEY",
)

completion = client.chat.completions.create(
    model="nvidia/nemotron-3-ultra-550b-a55b",
    messages=[{
        "role": "user",
        "content": "Hello, introduce yourself",
    }],
    max_tokens=500,
)
print(completion.choices[0].message.content)

Example configuration. Replace <gateway-host> and YOUR_API_KEY with your provisioned endpoint and key.

Separate reasoning output

reasoning_content keeps reasoning separate from the main response.

Native function calling

Automatic tool choice is enabled for agent workflows.

One protected gateway

TLS encryption, API-key authentication and per-project admission control through a single entry point.

NVIDIA Nemotron 3 Ultra
262K context · Open-weight MoE
TPSDEMO
120 t/sSustained tokens per second
AVERAGE LATENCYDEMO
2.4sIllustrative demo profile
SUCCESS RATEDEMO
99.9%Illustrative availability

REFERENCE PRICING

OPENROUTER REFERENCE

Input$0.50 / 1M
Output$2.20 / 1M

MODEL PROFILE

ARCHITECTUREACTIVECONTEXTCAPABILITIES
550B hybrid MoE55B262KTools + structured output
OpenRouter reference pricing; Aolani deployment pricing may differ.

MODEL

PROVIDERNVIDIA
TYPEOpen-weight MoE
CONTEXT262K
INPUTStext

Interactive open-model API playground with OpenAI-compatible examples.

SECURITY AND DATA

Control over the resources. Clarity on the data.

Dedicated deployment and gateway controls give production teams a clear operational boundary.

Dedicated resources

Requests and outputs stay within your exclusive cluster resources on dedicated endpoints, with whole-node isolation.

Never used for training

Customer requests and outputs are never used for model training.

Protected API access

TLS encryption, API-key authentication and per-project token-capacity admission control protect transport and access.

GETTING STARTED

A direct path to your production endpoint.

Bring your workload requirements. We’ll help define the model, endpoint and operating profile.

01

Scope your workload

Share model, context length, concurrency, latency and data-handling requirements.

02

Choose your endpoint

Select the Standard or Performance profile and shared or dedicated resources.

03

Integrate and validate

Connect with the OpenAI-compatible API and validate against your own production workload.

04

Operate and expand

Use managed monitoring and support, then adjust capacity as demand grows.

FAQ

Answers before you move to production.

Practical details on model access, performance, security and commercial options.

WHAT’S NEXT

The next layer of the platform.

Planned capabilities. Speak with Aolani about timing and availability.

PLANNED

Embedding and RAG stack

A complete retrieval-augmented delivery path, combining embedding models, storage and inference on one platform.

PLANNED

Multi-region deployment

Expanded regional deployment to meet customer compliance and latency requirements.

BUILD WITH AOLANI

Put your next model into production.

Discuss your model, traffic and data requirements with the Token Factory team. We’ll help you find the right endpoint and capacity.