What Is DMS Lab AI? OpenAI-Compatible Inference, Explained

Table of Contents

1. TL;DR
2. What is DMS Lab AI?
3. How the auto-router works
4. Why one endpoint instead of five integrations?
5. Who is this platform for?
6. What runs on this blog
7. Where to start
8. FAQ

TL;DR

DMS Lab AI is production AI infrastructure. You send one OpenAI-compatible request, and the endpoint profiles it for latency, precision, and context demands before routing it to the tier that handles that work best, from real-time speed to million-token reasoning. Your client keeps a single response contract from first token to final output. This post explains how that works and which of our guides to read first.

What is DMS Lab AI?

DMS Lab AI is an inference platform for teams shipping AI features into production. Instead of managing model choices, GPU capacity, and failover yourself, you integrate once against an OpenAI-compatible endpoint and let the platform make the per-request decisions.

The platform covers four layers: inference (the OpenAI-compatible API itself), knowledge (private context for your data), agents (tool use and multi-step execution), and automation (production workflows on top). The common thread is that every layer speaks the same request format your code already uses.

Concretely: if your application can call the OpenAI chat-completions API today, it can call DMS Lab AI by changing two environment variables. That is the entire migration path.

How the auto-router works

The core of the platform is dms-auto-router, a single model ID that stands in front of three inference tiers:

When a request arrives, the router reads its latency, precision, and context signals before any model answers, then selects the strongest available tier for that specific job. On our B200 fleet the flagship model pushes up to 720 tok/s aggregate with 85ms time-to-first-token — measured on our own fleet, not borrowed from a spec sheet. A one-line classification question still rides the fast lane. A forty-file repository trace gets the million-token window. Your client receives one consistent streaming response shape either way, so the routing decision never leaks into your integration code.

The workload changes; the endpoint does not. That is the contract.

Why one endpoint instead of five integrations?

Most teams end up maintaining a small matrix of provider SDKs, each with its own auth, rate limits, error semantics, and pricing model. Every new model launch becomes an integration project, and every migration risks breaking the features downstream.

Collapsing that matrix into one endpoint buys three things:

• One contract to test. Streaming, tools, and response shapes behave identically across tiers, so your test suite covers all of them at once.
• Per-request economics. Because routing happens per request rather than per contract, a product with mixed workloads, chat here, document analysis there, pays the appropriate rate for each instead of the highest common one.
• Room to re-platform. When a stronger open-weight model lands, it slots into a tier without any change on your side. Your code keeps working while the infrastructure underneath improves.

Who is this platform for?

Three workload archetypes get the most out of the platform, and each choice carries a concrete consequence:

• Teams shipping chat and copilot features. The Lite tier keeps interactive responses inside the 100ms first-token window users perceive as instant. Consequence: your chat UI feels native instead of laggy, at the lowest cost per token.
• Teams processing documents at scale. RAG pipelines and batch analysis on the Pro tier trade a slightly higher per-token rate for precision on reasoning-heavy extraction. Consequence: fewer silent wrong answers for users to catch in review, which is where the real cost of cheap inference lives.
• Teams building agents over large codebases or corpora. The Max tier's 1M-token window loads the whole context once instead of retrieving it in fragments. Consequence: multi-step agents stop losing the thread between turns.

And an honest boundary: if none of these describe you — for example, you need model training or fine-tuning, on-premises deployment, or non-text modalities — DMS Lab AI is not the right fit today. The platform serves text inference workloads, and pretending otherwise would only waste your evaluation time.

What runs on this blog

This blog is where we publish the evidence behind those claims. Three recurring formats:

• Performance benchmarks. Real numbers from our NVIDIA B200 and RTX 5090 GPU fleet, with methodology and reproduction steps, starting with Qwen 3.5 397B on B200.
• Integration guides. Step-by-step tutorials for Claude Code, Aider, Continue.dev, and the OpenAI SDK, starting with routing Claude Code through our endpoint.
• Model updates. What shipped, what changed, and what it means for workloads already in production, mirrored from the changelog.

Where to start

1. Benchmark first. Read the Qwen 3.5 397B benchmarks on B200 linked above, then reproduce the numbers with a free trial key. Trust scales with measurements you ran yourself.
2. Wire up your tools. The Claude Code integration takes about three minutes and two environment variables.
3. Go deeper. The API documentation covers streaming, tools, and batch endpoints, and the status page shows the live uptime of every component you would depend on.

FAQ

What is DMS Lab AI?
Production AI infrastructure: one OpenAI-compatible endpoint that profiles every request by latency, precision, and context, then routes it to the tier that handles it best. Built and operated by Vietnam DMS Company Limited.

Is DMS Lab AI compatible with the OpenAI SDK?
Yes. Any client that can call the OpenAI chat-completions API works by changing the base URL and API key. Migration is a two-variable change.

What models are available?
Qwen 3.5 397B for general reasoning (1M-token context in the Max tier), Qwen3 Coder for code generation, and a fast tier for latency-sensitive chat. See the API documentation for the current list.

Who is behind DMS Lab AI?
Vietnam DMS Company Limited, an infrastructure company operating its own B200 and RTX 5090 GPU fleet. Contact details and company information are on the contact page.

Can I try it before paying?
Yes. The free Pro trial includes 5 hours of inference time, enough to run our published benchmarks yourself.

---

Last updated: August 29, 2026. This article was drafted with AI assistance and reviewed by the DMS Lab team before publication.

Start building: Create a free account and send your first routed request in under five minutes.