Something is shifting in enterprise software — and it is not subtle. The companies that build the tools millions of people use for document management, productivity, and business processes are quietly adding a new requirement to their AI feature sets: the ability to run entirely on your own hardware, without sending a single byte to an external cloud service.
The Signal From This Week
Peter Steinberger, founder and CEO of Nutrient (formerly PSPDFKit), one of the most widely deployed document SDK vendors in Europe, made the direction clear in a post on X this week. Discussing progress in local AI tooling, he wrote that his team had spent significant time adding support for self-hosted models and LM Studio, and added: "we even have a maintainer from @ollama on the team. I love all the progress local models make!"
Nutrient's document processing SDK is embedded in applications across banking, legal, healthcare, and enterprise content management. When a vendor at that scale commits to bringing a core contributor from the Ollama project onto the team, it signals more than internal enthusiasm. It signals that enterprise buyers are starting to treat on-premise AI support as a product requirement, not a nice-to-have.
Why Enterprise Software Vendors Are Going Local
The pressure comes from multiple directions simultaneously:
GDPR and data residency: EU businesses in regulated industries face hard constraints on where data can be processed. Routing document content through an external cloud API can trigger data processing obligations — including Article 28 data processor agreements — that many organisations cannot accept. A local inference layer sidesteps the issue entirely without compromising functionality.
Internal security policies: Large enterprises routinely prohibit sending internal documents or proprietary code to external AI services. A locally-hosted model that never phones home removes this barrier without requiring policy exceptions or security review cycles.
Model quality thresholds crossed: In 2026, the gap between frontier cloud models and locally-hostable open-weight models has narrowed to the point where many business use cases are effectively indistinguishable. Community benchmarks report Llama 3.3 70B running at 30–45 tokens per second on a Mac Studio M4 Max — fast enough for synchronous document analysis workflows. Smaller models like Qwen 3.6-27B or Mistral Small 4 achieve 40–60+ tokens per second on standard workstation hardware, as reported by practitioners.
Total cost of ownership: A Mac Studio M4 Max sits at roughly €3,000–4,500. At typical cloud API rates for document-intensive workflows, that hardware investment can pay for itself within a few months of heavy use, after which the marginal cost of inference is essentially zero.
The Technical Stack: Ollama and LM Studio
Two tools are emerging as the defaults for enterprise local LLM deployment:
Ollama provides a server that exposes an OpenAI-compatible REST API, making it straightforward for applications to switch from cloud inference to local inference without major code changes. This compatibility is exactly why vendors like Nutrient choose it: the integration is often a configuration change, not a rewrite. An application already connected to a GPT-4 endpoint can point to a local Ollama instance by changing the endpoint URL.
LM Studio handles the desktop-side developer experience — with a graphical model manager and an integrated server that uses Apple's MLX framework on Apple Silicon. Community reports consistently show MLX-quantized models outperforming GGUF-quantized equivalents on Mac hardware, making LM Studio relevant for development teams doing local AI work on macOS before pushing to production Ollama instances.
Which Models Does the Community Recommend?
For 2026 enterprise use cases, current community recommendations cluster around three options:
- Qwen 3.6-35B-A3B: A mixture-of-experts architecture with only 3.5 billion parameters active per inference step from a total of 35 billion. Runs efficiently on 32–48 GB Apple Silicon systems, with quality approaching frontier models for many business tasks.
- Llama 3.3 70B: The high-capability standard for well-resourced hardware. 30–45 tok/s on Mac Studio M4 Max, as reported by community benchmarks.
- Mistral Small 4: The lightweight option for 16–32 GB machines, reported at 40–60+ tok/s.
A Concrete Integration Pattern
A practical enterprise integration looks like this: a document management system, a contract analysis tool, or a customer support platform connects to a locally running Ollama instance rather than an external cloud endpoint. The model processes the document, generates the summary or data extraction, and returns the result to the application. No content leaves the corporate network at any point. No log file in a third-party data centre retains a copy.
For application developers, this pattern is increasingly documented and tested in production. The Ollama OpenAI-compatible API means that existing integrations built for cloud models can often be adapted with minimal changes.
GDPR Compliance Benefits
Based on our reading of GDPR obligations, on-premise LLM integration provides meaningful compliance advantages:
No data processor agreement required (in most configurations): When data is not transferred to a third party, the processing relationship that triggers Article 28 obligations does not arise. Local processing stays within the organisation's own legal sphere.
Privacy by Design (Article 25 GDPR): On-premise architecture is a direct implementation of the data-protection-by-design principle. Processing capacity comes to the data rather than the data being sent to processing capacity.
EU AI Act scope: Internal tools using open-weight models generally face lighter regulatory requirements than cloud-based General-Purpose AI services offered to third parties. The compliance surface area shrinks considerably when there is no external service provider involved.
This is informational commentary based on our reading of the applicable framework. Individual legal situations vary — your data protection officer should assess your specific setup before drawing compliance conclusions.
What This Means for European SMBs
For small and medium-sized businesses across Europe, the enterprise software trend toward local LLMs creates a direct opportunity. Tools you already pay for may soon offer privacy-safe AI features without additional cloud contracts. But there is no need to wait for vendors to catch up:
Start a pilot: Identify one internal use case — document summarisation, email triage, contract clause extraction — and test it with a local model via Ollama before committing to a larger investment.
Evaluate software on local LLM support: Make on-premise inference a criterion in software evaluations. Vendors who provide it are signalling a meaningful commitment to user data control.
Check hardware support programs: Pan-EU digitisation support programs, as well as national schemes, may cover AI-capable workstations as part of broader digital investment initiatives. Consult your regional chamber of commerce or national authority for details.
More on how Freshlab helps European SMBs deploy local AI can be found on the Local AI page and the Kaira Toolkit. For businesses that treat data sovereignty as a strategic asset, the Data Sovereignty page offers further context. When you are ready to run a scoped pilot, the Pilot Project page explains how we structure engagements.
Next Step
Enterprise software is going local-first on AI. The tooling is mature, the models are capable, and the compliance arguments are on your side. The question is not whether your organisation will use local LLMs — it is whether you will be ahead of the shift or catching up to it.
Contact Freshlab to explore what an on-premise AI integration looks like for your specific use case and technology stack.