Platform
More than a router.
The roster is the storefront. Behind it is the practice — corpora, pipelines, finetunes, and the evaluation harness that keeps the whole stack honest.
Routing
One API. Seven names. No silent swaps.
Every request hits the same OpenAI-compatible endpoint and lands on a named model with a disclosed base. No mystery routing, no “powered by” fine print. The name you call is the model you get.
Curated context
A corpus we actually read.
We assemble, clean, and label domain corpora for a specific job — not a scrape of the open web. Provenance is preserved at ingest, so every passage can answer for where it came from.
Retrieval
The answer stays next to the evidence.
RAG pipelines wired into your stack, not a demo notebook. Retrieval is built to surface the source beside the claim, so the path from answer to evidence is one hop, and visible.
Finetuning
Your material. A named model.
We train on your corpus and serve the result through the same router, under a new name. Your model joins the roster with its base disclosed and its evaluation record attached.
Evaluation
Scored before it ships.
Every model and pipeline is stressed against task-specific harnesses: source-trace accuracy, claim support, calibration, retrieval precision. The harness exists to find failure before you do.
The API
One key. Every model.
The endpoint is OpenAI-compatible. Your tools, your agents, and your SDK already speak it. Pick a name and run.
curl https://api.trybased.ai/v1/chat/completions \
-H "Authorization: Bearer $BASED_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "based",
"messages": [{ "role": "user", "content": "Show your work." }]
}'Ways to engage.
API access
Request a key, point your stack, pick a name. The endpoint your tools already speak.
Scoped build
A corpus, a pipeline, or a finetune — scoped in a letter, built in your stack.
Embedded lab
Ongoing model work for teams that need a lab on call.
Asked, answered.
Who builds your models?
We build on frontier open bases — DeepSeek, GLM, Qwen, Kimi — and disclose the base on every model. Curation, specialization, and evaluation are the product. Pretending otherwise is not.
Why disclose base weights?
Because you cannot audit what you cannot name. The base tells you a model’s lineage, its likely ceiling, and its characteristic failure modes. We consider that the minimum you deserve.
Do you train on our data?
No. Client corpora are scoped, siloed, and used only for the engagement they were provided for. A finetune trained on your material is served under a new name, for you.
What does an engagement look like?
You write to us with the job. We reply with a scope, a timeline, and a fixed price. Short chains, no account managers, no discovery-call theater.
Can my agent use the API?
Yes. The endpoint is OpenAI-compatible: point it, pick a name, run. Agents read the roster the same way you do.
Access
Work with the lab.
API keys, scoped builds, and finetunes all start the same way: a letter. Tell us what you are building. We read every one.