You decide which models are allowed – large models via API or open models on your own servers. c:node picks the right one for each question. The check at the source stays the same for every model.
Anthropic
Via API in the EU region
OpenAI
Via API
Mistral AI, France
Via API · EU provider
Open model
On your servers · air-gapped too
Open model
On your servers · air-gapped tooYou don't have to take the biggest or most expensive AI model. Two things have changed: First, the freely available (“open”) models are now almost as good as the expensive, paid ones from the big US providers. Second — and more important for a business: for your specific task, a small model trained on your data beats the big all-rounders. And the more of your own data, the better it gets. That's exactly what c:node builds on: it uses open models of your choice and adds a knowledge store on top that grows with every case in your organisation — instead of locking you into a single provider.
A quick note on the percentages below: “accuracy of 90 %” means: out of 100 cases, the model classifies about 90 correctly. A “gap of 1.7 %” means: the freely available model is only 1.7 percentage points behind the best paid one — practically level.
For a long time, top performance was only available from the big, paid US models. That's no longer true. The Stanford AI Index 2025 measures how far the best freely available models trail the best paid ones — based on which answer people prefer in a direct comparison. This gap has practically shrunk away within a year:12
Only on the hardest tasks (such as programming or PhD-level specialist knowledge) are the paid models still a few points ahead. For the vast majority of everyday tasks the difference is small today — and freely available top models such as Kimi K3 sit right behind the market leaders at a fraction of the price.7
The decisive point for a business: you don't need a model that can chat about everything — you need one that reliably solves your task. That's exactly what scientific studies show again and again: a small, freely available model trained specifically on your domain data (experts call this “fine-tuning”) is as good as the big all-rounders or better — at a fraction of the size and cost. In the table, the small specialist model is on the left, the big all-rounder on the right; the number is the hit rate:
| The task | Small specialist model | Big all-rounder | Source |
|---|---|---|---|
| Explaining programming errors | Qwen-4B — practically level with GPT-4.1 | GPT-4.1 | arXiv 2507.05305 |
| Sorting topics correctly | Llama-2-7B — 86.4% correct | ChatGPT-4 — 83.4% | arXiv 2411.00890 |
| Assigning specialist categories | Llama-2-7B — 94.6% correct | Llama-2-70B (10× larger) — 84.9% | arXiv 2411.00890 |
| Classifying documents reliably | Qwen-3.5-27B (local) — 95.0% | Paid models — 75–80% | arXiv 2605.20368 |
Briefly explained: “94.6% correct” means the model was right in about 95 out of 100 cases. In the table, the small, specialised model almost always wins — even though the all-rounder is much larger and more expensive. The reason is the targeted training on the respective task.
Perhaps the most important finding: the model gets better the more of your own examples it receives. In one study, the hit rate of the same model rose from 84.8 % to 94.6 % — simply because it was given around 76,000 examples to learn from instead of 5,000.4 This is exactly where c:node comes in: c:node isn't a fixed, finished model, but a knowledge store that grows with every case in your organisation — the longer you work with it, the more precise and traceable the answers for your area become.
c:node brings exactly these findings together in three layers — separated and sealed off for each customer:
An open, pre-trained model provides language and general knowledge — interchangeable without changing the traceable logic. The foundation.
On top comes fine-tuning on your specialist tasks — exactly the step that brings small open models level with the big generalists (see §2 and §3).
Both ground every answer in your facts and firm rules — the neuro-symbolic part. That keeps the answer traceable – anything that can’t be evidenced is marked as open. The science behind it →
All three layers grow with every case, per tenant: more domain data sharpens the fine-tuning, every new piece of evidence extends the knowledge graph and ontology. The longer c:node works in your organisation, the more precise and traceable it becomes — isolated, on-prem if you wish, with data protection under the GDPR.
You don't depend on US providers for any of this. Germany's Aleph Alpha offers Pharia, a multilingual model (including German), and the complete stack including weights and runtime for on-premise operation — trained in line with EU and national law.5 c:node is model-agnostic: c:node runs on open models of your choice, interchangeable without changing the traceable logic — in the cloud, on your server or offline.
A fair comparison needs caution: on many well-known tests, the top models today all score above 90 % — the rest is hardly a difference any more, but measurement noise. A “number 1” headline therefore says little, because the test method alone often shifts the result. We deliberately show no day-to-day snapshot here, but the evidenced trends — and link the ongoing public leaderboards for you to check yourself:
The figures come from the Stanford AI Index 2025 and peer-reviewed studies (as of 2024–2026). Day-to-day model rankings change quickly — for snapshots, please use the linked live leaderboards. c:node is model-agnostic; the models named are examples, not an exclusive commitment.
Same product, same evidence – you choose where it runs.
Test instantly in the browser – free, with your domain.
Open the sandbox →Your own tenant in Frankfurt, separated from all others.
View security →c:node on your servers – offline if you wish.
View security →A short test run shows how c:node decides in your domain — traceably, with a source.