Managed model APIs
Stable, versioned endpoints for commercial inference, compatible with the SDKs your team already uses.
SiamGPT Token Factory
One key.
Production AI inference for Thailand
One key.
Every model.
Frontier, open-source, and Thai-native models through one OpenAI-compatible API. We operate the GPUs, routing, and metering in Thailand. Your team integrates one endpoint and ships.
OpenAI, Anthropic, and Gemini-compatible · Billed per token · Tokens are AI inference output, not cryptocurrency.
DIFF
app.py
ONE LINE CHANGED
1client = OpenAI(
2- base_url="https://api.openai.com/v1",
2+ base_url="https://api.siamgpt.com/v1",
3 api_key="sgt-•••a1",
4)
5
6client.chat.completions.create(
7 model="siamllm-th", # or claude-sonnet-5, gpt-5.5 …
8 messages=[{"role": "user", "content": "สรุปสัญญาฉบับนี้"}],
9)
One endpoint for
- SiamLLM
- OpenAI
- Anthropic
- DeepSeek
- Qwen
- Meta Llama
- Moonshot
- Z.ai
- Mistral
1API key per project, revocable in one click
3API styles: OpenAI, Anthropic, and Gemini
5Modalities: text, image, video, speech, embeddings
100%Of requests metered and attributed to a project
Managed inference
The token factory for commercial AI
SiamGPT Token Factory hosts and operates large language models for businesses through managed APIs and routing. You receive model outputs. We run the capacity, the serving stack, and the operations underneath.
Production token serving
We run the serving layer, capacity planning, observability, and reliability needed to produce tokens under real traffic.
Intelligent routing
Requests are weighed by health, latency, capacity, and cost, then routed with failover that switches in seconds.
Served in Thailand
Inference runs on NVIDIA GPUs operated in Thailand, so Thai data can stay in the country when your policy requires it.
Platform
Every way to buy inference, on one platform
Start pay-as-you-go, commit capacity once traffic is predictable, or isolate everything on hardware reserved for you. The API, keys, and bill stay the same.
Serverless inference
Call any catalog model on demand. No infrastructure to manage and no commitment.
Best for prototypes, new features, spiky traffic
Batch inference
Submit large offline jobs and collect results later, at a lower rate than real-time calls.
Best for document backfills, evaluations, embeddings
Provisioned throughput
Committed token capacity with a service level agreement, for production traffic that cannot queue.
Best for customer-facing assistants at steady volume
Dedicated endpoints
A model deployed on GPUs allocated to you alone, with predictable latency and your own configuration.
Best for latency-sensitive and custom models
Private deployment
The full serving stack inside your data center or an isolated enclave, operated by our team.
Best for government, finance, regulated data
Fine-tuning & evaluation
Adapt SiamLLM or open models on your data, and measure quality before anything reaches production.
Best for domain language, house style, accuracy targets
Governance
Who can call which model is decided by structure
Organizations, workspaces, and projects scope every key and every model. Permissions follow the structure, not individual people.
- Three-level hierarchyOrganization, workspace, and project each carry their own members, roles, and credit budgets.
- Model allowlistsA project can only call the model IDs enabled for it, checked before traffic reaches any provider.
- Scoped keysEach key belongs to one project. Revoke it and that project's access ends everywhere at once.
Access structureClick a level
Project: holds the API keys and the model allowlist. This is where access is actually enforced.
Model allowlist3 / 5 enabled
Illustration. The allowlist applies before traffic reaches a provider.
Request log
All2xxErrors
| 09:41:22 | siamllm-th | 200 | 12.4k tok | 3.10 cr |
| project support-bot · key sgt-•••a1 · workspace Customer Experience | ||||
| 09:41:20 | gpt-5.5 | 200 | 3.1k tok | 0.68 cr |
| 09:41:18 | kling-v3-pro | 200 | — | 42.00 cr |
| 09:41:15 | gemini-3.5-flash | 429 | — | 0.00 cr |
| 09:41:11 | nemo-retriever-embed | 200 | 0.8k tok | 0.04 cr |
Illustrative data. Failed requests are logged and not charged.
Metering & billing
Every token has an owner
Usage from every provider is logged per request and rolled into one bill. Credits flow down the hierarchy, so budgets are set before the spend, not reconciled after it.
- Per-request attributionModel, key, tokens, and credits on every call, traceable to a project and a department.
- Credit allocationCredits move from the organization to workspaces and projects, with hard limits where you need them.
- One invoiceSpend across every provider arrives as one statement, so finance deals with one counterparty.
A Monday morning
Developers, finance, and admins each see their own side
The same governance layer, opened by three people for three different reasons.
09:12 · DEVELOPER
Swap one base URL and a new model is live
Keep the OpenAI SDK. Scope comes from the project key, so there is no new authentication flow to build.
DIFFconfig.ts
1- model: "gpt-5.5",
1+ model: "siamllm-th",
2 baseURL: "https://api.siamgpt.com/v1",
09:40 · FINANCE
Open the invoice and every line already has a project
No more stitching together provider statements, billing cycles, and exchange rates before the month can close.
ProjectWorkspaceCredits
support-botCustomer Experience18,420
contract-reviewLegal9,115
catalog-imagesMarketing6,730
Total3 projects34,265
10:05 · ADMIN
Approve a model once, and the allowlist updates everywhere
Who can call which model is answered in one place, instead of logging in to each provider to check.
Organization allowlist4 / 5 enabled
Onboarding
From workload to tokens
Self-serve keys get you started in minutes. For committed and private capacity, a project runs in four steps.
Scope
We map the workload, target models, latency needs, throughput, and data residency.
Provision
We allocate and configure serving capacity for the agreed workload and delivery mode.
Serve
Your team integrates one API and receives model outputs, with keys and quotas per project.
Operate
We monitor routing, capacity, reliability, and token delivery as your demand changes.
Built in-house
SiamLLM, our Thai-native model
A large language model platform designed specifically for the Thai language. It sits in the catalog next to every other model, one model ID away.
- Native understanding of Thai linguistic structures and nuances
- Aware of Thai business etiquette, regulations, and social norms
- Served on NVIDIA GPUs in Thailand, and powering the SiamGPT assistant
model: "siamllm-th" Meet SiamLLM
Solutions
Built for your operating reality
Stability, cost, compliance, and delivery requirements differ by team. Token Factory adapts from standard API access to fully private deployment.
01 · INTERNET & SAAS
AI features that survive peak traffic
One model API integrated once, with failover and per-project quotas and billing.
02 · AI STARTUPS
From prototype to launch without a platform team
Switch models freely on one protocol and pay only for the tokens you use.
03 · GOVERNMENT & FINANCE
Compliance and audit as defaults
Private deployment with isolated compute in Thailand and full-chain audit trails.
04 · RESEARCH & EDUCATION
Compare many models, trace every run
Unified multi-model calls, elastic concurrency, and complete call logs.
05 · CHANNEL PARTNERS
Your own AI service under your brand
Your domain and tenant, separate keys, and revenue-split billing.
06 · CONTENT & MEDIA
Text, image, and video on one bill
Multimodal models behind one interface, with high-concurrency batch generation.
Start building on Token Factory
Tell us your workload, target models, and expected traffic. We will scope capacity and get your team an API key.