Keep your cloud's model catalog current.

A standing pod of LLM engineers and evaluators takes each new open-weight release, gets it serving on your stack, and checks it against the publisher's own results.

A new offer. It starts with one release, as a pilot.

Book a call Why hosts differ

For each release, the pod hands you

  • The model serving on your stack, with its configuration written down
  • A conformance report: chat format, tool calling, structured output, context length
  • Its scores beside the publisher's, and beside your previous version
  • A sign-off note and a working sample for your docs

The same weights score differently from host to host.

An open model's weights are identical on every cloud. What differs is how each host serves them: the serving build, the chat template and tool-call parser, quantization, and tuning for the host's hardware.

Each new release brings that work back. A catalog falls behind when no team owns it.

  • 93.3% to 36.7% was the range of scores for one open model on the same test across hosting providers, in a benchmark by Artificial Analysis in August 2025. Source
  • 80%, then 93.3% was one large cloud's score before and after it replaced an older serving build that ignored a setting. Source
  • Tool calling is where the gap shows first: one router now checks every tool call and routes around hosts with lower accuracy. Source

Serving software documents the same requirement: a chat template and the right tool-call parser for each model. vLLM documentation.

What the pod does when a release lands.

  1. Step 1

    Serve

    Package the release for your serving stack, with the chat template, tokenizer settings and tool-call parser the model expects.

  2. Step 2

    Conform

    Test chat format, tool calling, structured output, context length and sampling defaults. Where the publisher ships a compatibility test, run it.

  3. Step 3

    Measure

    Reproduce the publisher's benchmark scores on your endpoint, and compare with the version you serve today.

  4. Step 4

    Hand over

    A sign-off report, the configuration, and a working sample for your documentation.

Who is on the pod, and what stays with you.

The pod

  • LLM and machine-learning engineers who run model serving
  • Evaluation engineers who build and run the test suites
  • One lead who answers for the pod
  • Senior engineers in US time zones; roughly 1 in 40 pass our vetting

Your team keeps

  • Performance tuning for your own accelerators
  • Safety and legal sign-off on each model
  • Capacity and pricing decisions
  • The environment: the pod works inside it, under your access rules

How it starts.

Pilot
One release, one fixed scope and price. You judge the report before anything else is agreed.
After the pilot
A standing pod, run as staff augmentation on a monthly fee set by team size.
Contract with
Hire Llama Inc., the Florida company that operates Awana.
Access
Inside your environment, through your SSO. Each engineer signs your NDA first.
Not included
Training or fine-tuning. That is a different job: AI-training teams.
Security review
The vendor facts are on one page: Security and contracting.

A new offer, built from two things we already do.

We have not yet done this for a cloud provider. The pilot exists so you can judge it on one release.

  • We recruit LLM and machine-learning engineers for US companies. LLM engineers · ML engineers
  • We staff evaluation pods for AI labs: test sets, blind grading, agreement tracked on every batch. AI-training teams
  • 50 engineers in three months on AI-training work for a frontier AI lab, whose name we do not publish.
  • 500+ engineers placed with US companies since 2020; 98% stay past the three-month guarantee.

Who you talk to.

Calvin Sedao

Calvin Sedao

CEO and Co-Founder, Seattle

Founded Awana in 2020 after building a point-of-sale company with a Venezuelan engineering team.

Javier Dutan

Javier Dutan

Co-Founder and COO, New York

Former White House US Digital Service.

One lead answers for the pod. Awana is United States-led, and its leadership holds an active US security clearance. Calls book on Calvin's calendar.

Questions a platform team asks.

Do you train or fine-tune the model?

No. An open model's weights are the same on every cloud. This is serving and verification work: getting each release running correctly on your stack and proving it. Tuning a model on your own documentation and APIs is a different job, covered by our AI-training teams.

Why would the same model score differently on our cloud?

Because of how it is served. An older serving build, a chat template or tool-call parser that does not match the conventions the model was trained on, a quantization choice, or tuning for the host's hardware can each move the result. In one public benchmark of a single open model across hosts, scores ran from 93.3% down to 36.7%.

Which models does the pod work on?

The open-weight families you carry or want to carry. You choose the releases and the order.

Do you tune performance for our hardware?

No. Tuning for your own accelerators stays with your performance team. The pod gets the release serving correctly, tests it and reports what it measures.

How fast is a release turned around?

We set that target with you in the pilot. We do not publish a number we have not earned.

How is it priced?

The pilot is one release at a fixed scope and price. After it, a standing pod runs as staff augmentation, on a monthly fee set by team size. You get a quote after the first call.

Have you done this for a cloud provider before?

Not yet. Awana recruits LLM and machine-learning engineers for US companies and staffs evaluation pods for AI labs. This pod puts the two together. The pilot exists so you can judge it on one release.

Tell us which release is stuck.

Thirty minutes on your catalog, your serving stack and where the time goes when a new model comes out.

Book a call Security and contracting

Related: AI-training teams · AI red teaming and evaluation · Staff augmentation