For work where anonymization makes no sense – or simply isn't needed.
anymize Waterfall 2 and anymize Fountain 2 are the second generation of our own AI models. Waterfall 2 is based on GLM 5.3 Flash and is the new default model in chat and in the API. Fountain 2 is based on NVIDIA Nemotron 3.5 Lightning and is the fast model for agent tasks and connected data. Both handle up to 1M tokens of context. Available in chat and via API, with DPA and the consent statement under § 43e BRAO for German lawyers (equivalent professional-confidentiality frameworks apply across most European jurisdictions). For code, Excel, SQL and structured data.
Why in-house models?
For sensitive text the answer is clear: anonymize before sending to a frontier model. That's our second track – GDPR-compliant and powerful. But there are tasks where that route doesn't work:
loses its structure when variable names, class identifiers or function signatures are swapped for placeholders.
break when cell references or column names are anonymized.
become inconsistent when keys disappear.
lose their semantics when table or column names turn into tokens.
For these cases you need a model that can work with the real structure of your data.
That's exactly what anymize Waterfall 2 and anymize Fountain 2 are for.
The models
Two models, two characters. Waterfall 2 is the new default model for demanding tasks, Fountain 2 the fast model for agent tasks and connected data. Both handle up to 1M tokens of context.
Default model
Our new default model for fast reasoning, strong coding and reliable tool use.
| Property | Value |
|---|---|
| Version | 2 · since August 2026 |
| Category | Deep |
| Base | GLM 5.3 Flash |
| Context window | up to 1,048,576 tokens (1M) |
| Reasoning | yes, thinks before answering |
| Available | default model in chat and in the API |
| Use | complex analyses, coding, multi-step tasks with tool use, long documents |
Performance
Our fast model for agent tasks and working with connected data.
| Property | Value |
|---|---|
| Version | 2 · since August 2026 |
| Category | Swift |
| Base | NVIDIA Nemotron 3.5 Lightning |
| Context window | up to 1,048,576 tokens (1M) |
| Strengths | short response times, tool use |
| Available | in chat and in the API |
| Use | fast agent tasks, working with connectors, long-running workflows, structured data work |
Both together
One stack, two characters. Waterfall 2 thinks thoroughly, Fountain 2 delivers fast. Pick per question or let anymize recommend.
Note on currency
Since August 2026, Waterfall and Fountain run in their second generation. Fountain 1.0 (Smart) remains selectable in chat, Waterfall 1.0 remains available for existing integrations.
German-language fine-tuning
GLM 5.3 Flash and NVIDIA Nemotron are among the strongest open models in the world. But: open-source models are optimized on English-language benchmarks. Continental European legal, medical and commercial vocabulary – and German contract, tax and labor law in particular – are underrepresented in the base models.
That's why we additionally train Waterfall and Fountain on a carefully curated German-language professional dataset:
laws, judgments, pleadings, standard contract templates
guidelines, specialist publications, treatment documentation
annual reports, economic reports, tax literature, business administration
All publicly accessible domain data, carefully curated – no synthetic data. The result: for German-language professional contexts (useful across Germany, Austria, Switzerland and any European team working with German-language material), our models are often more competitive than pure benchmark numbers suggest.
Open-source models are typically one version behind the proprietary frontier models from OpenAI, Anthropic or Google. Close – but not state of the art. Our edge: the German-language professional fine-tuning. For anyone analyzing German contracts, drafting German pleadings, structuring German findings, our answer is often the better one compared to an English-dominated frontier model.
And if you need the absolute top quality only frontier models deliver: that's what Track 1 is for – GPT, Claude, Gemini and friends, GDPR-compliant through automatic anonymization.
Where they fit
Six scenarios from practice – with model recommendation:
Variable names, function signatures, comments stay intact – no anonymization
Cell references, formulas, sheet names stay semantically untouched
Table/column names stay readable – test with real data
Strong German-language legal vocabulary + reasoning + long documents
Strong German-language medical vocabulary + clear structure
Long documents, deep reasoning chains, up to 1M tokens of context
When Track 1 (international models) is the better fit: general-knowledge questions, creative text work, multimodal tasks, absolute frontier quality. That's where GPT, Claude, Gemini and friends play their strengths – and anymize automatically anonymizes for you before sending.
Access & legal framework
Waterfall and Fountain are available in two ways:
The API lets you decide per request: with or without anonymization parameter. With Waterfall and Fountain, anonymization is not necessary – but of course possible if you want it as a second security layer.
Use is governed by the contracts signed with you, and they apply directly:
anymize Spring · In preparation
anymize Spring is our third in-house model – currently in development. Significantly smaller than Waterfall and Fountain, but lightning-fast and resource-frugal. Two distinctives:
for real-time classification, simple text transformations, autocompletion, quick previews.
Spring is offered as a model that can run in your own infrastructure. For organizations that have to host model inference themselves (banks, insurers, public sector with special requirements).
Waterfall 2 and Fountain 2 run as a service through anymize. Spring extends our portfolio with the on-premise option for customers with maximum data-sovereignty requirements.
Available soon.
Who benefits
Code work with real structures – embeddable into your own tools via API
Excel/CSV/SQL analyses on company data, structured extraction, JSON transformations
Clean contract chain, one contract partner, predictable compliance
No extra DPA, automatic DPA, consent statement under § 43e BRAO (for German lawyers; equivalent frameworks across EU), audit trail
Strong German-language legal vocabulary after fine-tuning (useful for teams working with DACH-region material), no anonymization needed for internal work
Strong German-language medical vocabulary, DPA + confidentiality included
And: every organization that wants to use AI productively, but wants to clear up the question "where does our data go?" once and for all.
Frequently asked questions
Waterfall 2 is our new default model, based on GLM 5.3 Flash: fast reasoning, strong coding and reliable tool use. Fountain 2 is our fast model, based on NVIDIA Nemotron 3.5 Lightning: short response times, well suited to agent tasks and working with connected data. Both handle up to 1M tokens of context. Rule of thumb: Waterfall 2 for thorough work, Fountain 2 for speed.
All models. All features. No credit card.
Rather watch first? Free webinar – anonymization live
We stand behind anymize. And we know – when an AI tool touches client, patient or employee data, a demo video isn't enough. That's why we give you 14 days of full access – all models, all features, no credit card. Enough time to be certain, before you trust us.
Your AI workplace awaits.