Ollama has been the default starting point for running large language models locally — Llama 3.3, Qwen2.5, Gemma 3, DeepSeek-V3. Until recently, it relied on llama.cpp as its primary inference engine across all platforms. That has now changed in a meaningful way: Ollama has officially switched to MLX as its primary backend for Apple Silicon.
The official Ollama account announced on X that "Ollama is now updated to run the fastest on Apple silicon, powered by MLX, Apple's machine learning framework" (Ollama on X). For anyone running a local LLM stack on a Mac — whether a MacBook Pro, Mac Mini, or Mac Studio — this is the most significant performance architecture change in the tool's history.
Why MLX outperforms llama.cpp on Apple Silicon
MLX is Apple's own open-source machine learning framework, built to leverage the unified memory architecture of M-series chips. Where a conventional GPU setup maintains separate pools of CPU RAM and VRAM, forcing data copies between them, Apple Silicon uses a single memory pool that all compute units — CPU cores, GPU cores, and the Neural Engine — access simultaneously.
This difference is not a minor implementation detail. Transformer model inference is memory-bandwidth-bound: the speed at which weights and activations move through memory determines how many tokens per second the system can generate. By eliminating data copies between memory regions, MLX removes a fundamental bottleneck that llama.cpp, optimised for cross-platform compatibility, could not fully address.
The practical result: the same Mac hardware runs local LLMs measurably faster after this update, with no configuration changes required from the user.
What speeds to expect
Community benchmarks from practitioners reporting on the new MLX backend — not Freshlab measurements — indicate the following performance ranges:
- Mac Mini M4 Pro (48 GB): 7B models (Qwen2.5-7B, Llama 3.2) in the range of 60–90 tok/s
- Mac Studio M3 Max (96 GB): 13B models around 40–60 tok/s; 30B models around 25–38 tok/s
- Mac Studio M3 Ultra (192 GB): 70B models (Llama 3.3 70B) reported at 20–35 tok/s
Actual performance depends on quantisation level, context window size, and concurrent load. The key point for existing Ollama users: the improvement is automatic upon updating to the latest release. No backend switching, no configuration editing — Ollama selects the MLX path transparently on supported Apple Silicon hardware.
Coding agents are the headline beneficiary
Ollama's announcement specifically highlights coding agents as a primary beneficiary of the MLX switch, naming tools like OpenCode as examples. The reasoning is straightforward: a coding agent generates, evaluates, and iterates on code fragments in rapid succession. Each additional second of latency disrupts the developer's flow. A faster inference backend translates directly into a more responsive coding assistant.
This matters beyond convenience. For software teams that handle proprietary code, architectural decisions, or regulated data — and want to avoid routing any of it through external AI endpoints — running a local coding agent is the only option that satisfies data sovereignty requirements. With Ollama's MLX backend on a Mac Studio, the entire pipeline runs on the local machine. No source code, no comments, no internal identifiers leave the premises.
For an overview of how local AI integrates with team workflows while satisfying GDPR requirements, see our resource page.
Mac clusters: the emerging architecture
A discussion gaining momentum in the developer community: running multiple Mac Mini M4 or Mac Studio units as a distributed inference cluster. Practitioners on X are exploring this as a logical extension of Apple Silicon's unified memory architecture combined with emerging MLX-based cluster libraries.
The theoretical case is compelling. Two Mac Studio M3 Ultra units, each with 192 GB of unified memory, could in a distributed configuration handle models exceeding 300 billion parameters — a capability tier that has historically required enterprise-grade NVIDIA accelerators at significantly higher cost and power draw.
This remains an area of active experimentation rather than fully production-ready tooling. But the trajectory is notable: local AI infrastructure on Apple Silicon is scaling upward in capability without a proportional increase in complexity or cost.
Compliance implications for European businesses
For organisations operating under GDPR and looking ahead to the EU AI Act, Ollama's evolution has direct practical relevance. The fundamental advantage of an on-premise local LLM deployment is that it satisfies data sovereignty requirements by architecture: no personal data, no business data, no proprietary information is transmitted to a third-party inference provider.
Based on our reading of EU AI Act Article 26, deployers of AI systems in scope face documentation and monitoring obligations from August 2026. A local Ollama deployment gives operators full, auditable visibility into model versions, inference parameters, and data flows — which is structurally easier to document than a cloud API dependency where model versions and data handling practices can change without notice.
For a practical walkthrough of what a GDPR-compliant local AI stack looks like, visit our data sovereignty page. For an overview of EU AI Act deployer obligations and timelines, see local-ai.html.
Choosing the right hardware today
The Ollama MLX update reinforces Apple Silicon as the most practical platform for local LLM deployment in a European SMB context. The combination of energy efficiency, unified memory, and now a first-class inference framework makes the value proposition clear. Practical guidance by tier:
- Entry level (7B–13B models, single user): Mac Mini M4 Pro with 48 GB — covers the majority of SMB use cases at a compelling price point
- Mid-range (13B–30B models, light multi-user): Mac Studio M3 Max with 96 GB — balanced performance and cost
- High performance (70B models, multi-user team deployment): Mac Studio M3 Ultra with 192 GB — the current ceiling of cost-effective local LLM hardware
For teams with CUDA-dependent workloads or requirements beyond what Apple Silicon can address, NVIDIA options like the DGX Spark (GB10) remain relevant. But for Ollama-based stacks specifically, the MLX update makes the Apple Silicon path stronger than it has ever been.
Next steps
Existing Ollama users on Apple Silicon: update to the latest release. The MLX backend activates automatically for supported models. The upgrade takes minutes; the benefit is immediate.
Teams evaluating a local LLM deployment for the first time: the barrier has just dropped. A modern Mac Studio with Ollama running MLX-native models is a production-capable private AI stack, not a hobbyist project.
If your organisation wants a structured assessment — which models fit your use cases, what hardware matches your budget, how integration with existing systems works — that is exactly what our pilot engagements cover. Start with a no-commitment scoping call at /pilotproject.html.