Why does Guztia publish which tools it uses?

Because the stack is not a secret. LiteLLM, Langfuse, vLLM and the open-weight models are well-known open-source software anyone can inspect. The value is in knowing which workloads can move without anyone noticing, how to size committed capacity against demand that has not settled, and what the right mix is for a given organisation. That judgement is the product. The complete proposed design for your organisation is shared under NDA before any commitment.

Could we build this ourselves?

Yes. If you have a platform team with spare capacity and an appetite for owning it, you probably should. Guztia is a reasonable choice when you do not, or when that team's time is worth more on your product. What organisations underestimate is not the build — it is the eighteen months afterwards, during which it has to be kept current and is nobody's job.

Is any of this proprietary to Guztia?

No. The components are established, widely used, open-source software that Guztia did not write and does not licence to clients. Everything runs under accounts in the client's own name. If the arrangement ends it keeps working, and another firm or an internal team can take it over.

Where do open-weight models still fall short?

Long-horizon agent tasks running many steps with tool use, genuinely novel reasoning problems, quality at the far end of very long context windows, low-resource languages, and the harder multimodal document tasks. For the bulk of ordinary enterprise work — classification, extraction, summarisation, routine code, internal question answering — the difference in outcome is usually not detectable.

How It Works — LiteLLM, Langfuse, vLLM

01 The shape of it

Fig. 01 — The gateway is the seam: left of it nothing changes, right of it everything can.

02 Four layers, all open source, all yours

You'll own every layer.

Typical technologies shown. Final selection depends on your workload, existing infrastructure and operational preferences — we propose the specific design under NDA before any commitment.

01 The gateway LiteLLM Proxy A single endpoint every team talks to instead of talking to a supplier directly. Routing, failover, spending limits per team, and the point at which switching supplier stops being an engineering programme. MIT

02 The record Langfuse Every request logged with who asked, what answered, how long it took and what it cost. This is what turns one unexplained invoice into a breakdown you can charge back and defend. MIT

03 The serving layer vLLM · SGLang What turns capacity you have bought into work delivered. How well this is configured decides your cost per unit of work, and it is where most self-assembled attempts quietly lose their savings. Apache-2.0

04 The models Qwen · DeepSeek · Llama · Mistral Open-weight models, run on infrastructure in your name. We benchmark candidates against your evals and route accordingly — good enough for the bulk of ordinary business work, and closing the gap on harder tasks with every release cycle. Varies

03 Why open source, specifically

Because the point is that you can fire us.

A managed service only its vendor can operate is not a solution to supplier lock-in. It is a smaller supplier with the same problem.

Everything we deploy is permissively licensed, runs in your accounts, and is documented publicly by people who are not us. That is not altruism — it is the only version of this business we think survives contact with a competent CTO, and it is why we charge for operating the arrangement rather than for access to it.

Not in the arrangement

Any Guztia-built software, any proprietary control layer, any margin on your usage, any hosted service of ours that your traffic must pass through.

You hold

The cloud accounts, the infrastructure code, the model weights, the usage records, the documentation.

We hold

A monthly invoice and the pager.

Leaving

Thirty days' notice. Nothing to extract, nothing to rebuild, no data to get back.

04 Where open weights still lose

We are not going to tell you the gap has closed.

It has narrowed a great deal. For the bulk of enterprise work — classification, extraction, summarisation, internal question answering, routine code — a good open-weight model is not distinguishable from a premium API in outcome, while costing a fraction. That is most of your volume, and it is the whole basis of the argument.

It is not all of it. There remain tasks where the premium models are simply better, and pretending otherwise would cost you more than you saved. That is exactly why every arrangement we build keeps a capped route to a premium vendor rather than burning the bridge.

Which of your workloads fall on which side is an empirical question about your work, not a matter of opinion — and answering it is the first thing an assessment does.

Long-horizon agents

Tasks running many steps with tool use and self-correction. Premium models hold together noticeably longer before drifting.

Genuinely novel reasoning

Unfamiliar problems rather than variations on well-covered ground. The gap is real and shows up clearly in testing.

Very long context

Open models advertise large windows; quality at the far end of them is a separate question. Test at the length you will actually use.

Low-resource languages

Coverage varies sharply. If you operate in one, it has to be tested specifically — a headline multilingual claim will not answer you.

Harder multimodal work

Open vision-language models are improving quickly but premium APIs remain ahead on difficult document and image reasoning.

THE RECOMMENDATION

Route by difficulty, not by loyalty

Bulk volume goes to capacity you control. Harder work goes to a larger model you also control. What neither can do well goes to a premium vendor, under a hard ceiling — and every one of those requests is recorded, so you can see precisely what the escape hatch costs and whether it is shrinking. In the arrangements we run, it shrinks.

What that does to the bill

How It Works

You'll own every layer.

Because the point is that you can fire us.

We are not going to tell you the gap has closed.

Route by difficulty, not by loyalty

See what your AI actually costs.