Serverless inference
The fastest way to run catalog models on demand. No infrastructure to manage and no long-term commitment. Capacity scales with your traffic automatically.
Best for prototypes, new features, spiky traffic
Token Factory › Platform
Platform
From GPU to token, operated for you
Inference, model shaping, governance, and billing on one platform. Pick the delivery mode that fits the workload today, and move between them without changing your integration.
Inference
Five ways to run a model
All five share the same API, keys, allowlists, and bill. What changes is how capacity is allocated underneath.
Batch inference
Process large workloads asynchronously at a lower rate. Submit a file of requests, and collect results when the job finishes.
Best for document backfills, classification, embeddings, evaluations
Provisioned throughput
Committed inference capacity with token-based pricing, reserved throughput, and a service level agreement. Drop-in compatible with serverless calls.
Best for production assistants at predictable volume
Dedicated endpoints
Deploy a model on GPUs allocated to you alone. Control batch size, context length, and scaling, with latency that does not depend on other tenants.
Best for latency-sensitive products, fine-tuned models
Private deployment
Local deployment and private model hosting inside your facility or an isolated enclave, with dedicated operations from our team.
Best for government, finance, healthcare, sovereign data
GPU capacity
Training, inference, and batch capacity on NVIDIA GPUs for teams that need raw compute alongside managed endpoints.
Best for custom training runs and research workloads
Model shaping
Make a model yours, then prove it works
Fine-tune SiamLLM or open-weight models on your own data. Improve accuracy, reduce hallucinations, and control behavior, without managing training infrastructure.
Fine-tuning
Supervised and preference tuning on your corpus. The resulting model deploys to a dedicated endpoint behind the same API.
Evaluations
Measure model quality against your own task set, compare candidates side by side, and set a bar before production.
Optimization
Quantization, caching, and acceleration tuned to your scenarios, improving quality while cutting long-run cost.
Delivery stack
Seven capabilities in three layers
Resources keep it stable, the platform keeps it governed, and a single API keeps it simple.
Access
API
High-availability inference API
One entry point for chat, generation, retrieval, function calling, and embeddings, in OpenAI, Anthropic, and Gemini request styles. Streaming by default.
↑
Platform
Model hub
Unified model catalog
Language, embedding, image, video, and speech models behind one protocol. Switch with a model ID, not a migration.
Control
Organizations & projects
Isolated workspaces with their own keys, roles, allowlists, quotas, logs, and bills.
Billing
Token metering
Usage-based pricing with tiered discounts and project-level split billing. Every call traceable to a cost.
Optimize
Model operations
Fine-tuning, quantization, caching, and acceleration for the scenarios that matter to you.
↑
Foundation
Private
Private deployment
Local hosting for data security, compliance, and dedicated performance, with dedicated operations.
Resource
NVIDIA GPU pools
Serving, training, and batch capacity in Thailand, with elastic expansion as demand grows.
Delivery modes
Start small, commit when it makes sense
Pricing depends on model, volume, and delivery mode. Our team will quote against your actual workload.
Pay as you go
Top up credits and call any enabled model. Pay only for tokens used.
- Serverless and batch inference
- Full model catalog
- Organizations, projects, scoped keys
- Per-request logs and usage dashboard
Provisioned
Reserved token capacity with a service level agreement for steady traffic.
- Everything in pay as you go
- Committed throughput with SLA
- Tiered volume discounts
- Named technical contact, 24/7 support
Dedicated & private
Single-tenant endpoints or the full stack inside your own environment.
- Dedicated GPUs or on-premises deployment
- Data residency in Thailand
- Fine-tuning and evaluation programs
- Brand partnership under your own domain
Architecture
The path a request takes
From application integration down to model and compute resources, every layer carries its own security, scheduling, governance, and failover.
Application
Your product
↓
Access
Unified access
↓
Governance
Security & governance
↓
Routing
Intelligent scheduling
↓
Resource
Models & compute
Security matrix
Transport encryptionAccess whitelistsTenant data isolationFull-chain loggingMulti-node disaster recovery
Service level
More than a connection. A service level.
01
Multi-dimensional routing
Load, latency, capacity, health, and cost weighed together, with failover that switches in seconds.
02
Enterprise governance
Tenants, permissions, keys, quotas, audit trails, billing, and risk control, end to end.
03
Full stack, GPU to token
Infrastructure and model operations optimized as one path, so customization and response are fast.
04
One protocol, every vendor
Standardized calls across providers remove duplicate development and migration cost.
05
Three delivery modes
Public cloud, brand partnership under your own domain, or fully private deployment.
06
Cost-aware scheduling
Pay for what you use, while scheduling keeps reducing long-run spend.
Infrastructure
NVIDIA Technology Implementation
Token Factory runs on the same infrastructure plan as SiamLLM, so the capacity you call through the API is capacity we operate ourselves.
Computing Infrastructure
We plan to deploy NVIDIA DGX B200 for training and serving large AI models. Our products run on NVIDIA GPUs.
AI Development Frameworks
- For data ingestion, we implement the NVIDIA NeMo Retriever, followed by a data curation process. We then utilize the synthetic data generation pipelines from NVIDIA NeMo Curator.
- The Query Orchestrator manages initial processing, creates pre-filters, and builds context through vector search. It then leverages NVIDIA NeMo Retriever and NIM microservices – including Deepseek R1, Riva NMT, Llama 3.1, NeMo Retriever reranking, and NeMo Retriever embedding NIM – to produce accurate and contextually relevant responses.
Tell us about your workload
Share your target models, expected traffic, and data residency needs. We will recommend a delivery mode and quote against it.