The most widely used local LLM runner just received a platform-level upgrade. The official Ollama account on X announced that the tool is "now updated to run the fastest on Apple silicon, powered by MLX, Apple's machine learning framework." For European businesses running Mac hardware, this is directly actionable.
This is not an incremental patch. It is a change to the inference engine underneath every model you run through Ollama on macOS. Every workflow, every model, every use case gets faster — without any configuration changes on your end.
What Changed: From llama.cpp to MLX
Until this update, Ollama ran inference on macOS through llama.cpp — a mature, cross-platform C++ library that works identically on Windows, Linux, and macOS. Cross-platform consistency is valuable, but it means llama.cpp cannot take full advantage of hardware-specific capabilities.
MLX is different. It is Apple's own tensor computation framework, built from scratch for Apple Silicon's unified memory architecture. On M2, M3, and M4 chips, the CPU and GPU share a single contiguous memory pool — no data needs to be copied from system RAM to a discrete GPU's VRAM before inference can start. MLX was designed to exploit this directly.
The result is that operations which previously required multiple memory transfers now happen in-place. For large language models, which are fundamentally large matrix operations, this has a measurable effect on throughput, latency, and energy efficiency — particularly for longer context windows and larger model sizes.
Real-World Performance: What Practitioners Report
The impact is not purely theoretical. Practitioners across X and community benchmarking threads report noticeably faster token generation after the update. Based on measurements shared by the community, throughput gains of 20–50% over the previous llama.cpp backend have been reported on equivalent hardware, depending on the model size and quantization level. These are community-reported figures, not Freshlab-verified benchmarks — your actual results will depend on your specific configuration.
To put that in concrete terms:
- Mac Mini M4 Pro, 64 GB — 32B-parameter models reported at 25–40 tok/s, comfortable for interactive use and single-user workflows
- Mac Studio M3 Ultra, 192 GB — 70B-parameter models reported at 15–25 tok/s, viable for production document processing and multi-user setups
- MacBook Pro M4 Max, 128 GB — strong for developers who need a portable local LLM with no internet dependency
These are the same machines European SMBs already purchase for general office work. The marginal cost of running local AI on hardware you already own is close to zero — a significant contrast to pay-per-token cloud API pricing at scale.
Every Model in Ollama's Library Benefits
Because the MLX upgrade is an engine change rather than a model-specific optimisation, every model available through Ollama inherits the improvement:
- Llama 3.3 70B — Meta's latest open model with strong instruction following and multilingual output, including German and Spanish
- Qwen2.5 32B — Alibaba's multilingual flagship with particularly good quality in European languages; practitioners report it handles German formal register well
- DeepSeek-V3 — strong on structured reasoning, code generation, and long-document analysis
- Gemma 4 27B — Google's instruction-tuned model with native function calling, well-suited to agentic workflows
Model selection depends on your use case and hardware. For general-purpose business tasks — summarisation, drafting, classification — a 14B or 32B model often gives a better speed-quality trade-off than a 70B model on the same hardware.
GDPR: Nothing Leaves Your Mac
This is the aspect that matters most for European businesses operating under GDPR. When you run a model through Ollama on your own hardware, every token — input prompt and generated output — stays on that machine. The model weights load into local memory. There is no outbound API call, no telemetry endpoint, no vendor logging.
This matters because GDPR Article 32 requires "appropriate technical and organisational measures" to protect personal data. A local inference stack where data physically cannot leave your premises is a genuinely strong technical control — not just a contractual one.
For teams handling legal correspondence, HR documentation, medical records, or financial analysis, this means you can run AI-assisted workflows against sensitive content without signing a Data Processing Agreement with a third-party API provider, and without relying on that provider's privacy policy holding under future regulatory scrutiny.
The Ollama + MLX stack gives you speed and data sovereignty simultaneously — which was previously a harder trade-off to make.
Setting Up: Nothing Changes Except the Speed
If you already have Ollama installed on an Apple Silicon Mac, updating is sufficient. The MLX backend activates automatically — no configuration file changes, no additional framework installation.
# Update Ollama, then pull and run your model
ollama pull qwen2.5:32b
ollama run qwen2.5:32bIf you are new to Ollama, installation takes around five minutes. The Ollama documentation covers the setup process in full. Hardware sizing guidance from the community suggests 64 GB of unified memory as a practical starting point for business use, enabling 32B models at usable speeds with room for the operating system and other applications.
Developer Tooling: Xcode + Local Ollama
One downstream consequence of the MLX engine upgrade is that Xcode's Apple Intelligence integration with Ollama becomes meaningfully faster. Developer Anders Brownworth noted on X that Xcode's Apple Intelligence can be configured with a local LLM via Ollama for private coding assistance — no internet connection required. With the MLX backend in place, this integration is noticeably more responsive.
For development teams building iOS or macOS applications, this means AI-assisted code completion that runs entirely on local hardware — a meaningful consideration for teams working with proprietary codebases or client-side logic where sending code snippets to a cloud provider is undesirable.
The Ollama project's official announcement also cited Claude Code and OpenCode as tools that can benefit from the MLX upgrade, reflecting the broader trend of development tooling moving toward local AI backends where privacy requirements are tight.
The Funding Angle for European SMBs
Mac hardware investment is real. European funding programmes can offset a significant portion of the cost.
Germany: The BAFA "Digital Jetzt" programme and KfW digitisation loans cover investments in on-premise IT infrastructure. A local AI stack running on Mac Studio hardware can qualify as a digitisation investment under several programme categories, based on our reading of current scheme guidelines. Consult your Steuerberater for individual eligibility — the key documentation requirement is demonstrating the business use case for the hardware.
Spain: The Kit Digital subsidy covers AI and digital tool adoption for SMBs with 3–49 employees. Local AI infrastructure may qualify under the "Inteligencia Artificial y Analítica" category. Freshlab is an accredited Kit Digital provider — see our detailed guide at /kit-digital.html for current eligibility criteria and application steps.
EU-wide: InvestEU and national EIB-channelled programmes fund SMB digitisation. Check with your national development bank for active schemes in your country.
Building on the Stack
Faster local inference is an enabling capability. The business value comes from the workflows you build on top of it. Use cases where the Ollama + MLX stack pays for itself quickly:
- Document intelligence: Automated summarisation of contracts, invoices, and regulatory filings against a local knowledge base, without sending content to a cloud API
- Customer service drafting: A local assistant that generates draft replies in your brand voice, reviews them for compliance language, and flags escalations — all within your premises
- Internal search and Q&A: A retrieval-augmented generation setup over your internal documentation, giving staff accurate answers without exposing proprietary content
- Code review and generation: Internal development teams using a local model as a code review assistant, particularly useful where client NDAs restrict use of cloud coding tools
For a structured assessment of which of these use cases fits your current data and infrastructure, Freshlab offers pilot projects that run on your actual documents over two weeks. We also offer training for technical teams on managing Ollama-based local AI stacks in production.
For specific questions about hardware sizing, model selection, or GDPR documentation requirements for your deployment, contact us.