Gemma 4 Runs 90% Faster on Apple Silicon in Ollama

Local AI Just Got Faster Where It Counts — On the Laptop in Front of You

The best AI upgrades are the ones you feel instantly, and this is one of them. With Ollama v0.32.0, released July 11, 2026, Google's open Gemma 4 models now generate tokens nearly 90% faster on Apple Silicon. That is not a lab number — it is the difference between a local model that feels like a demo and one that feels like a tool you reach for all day. For the growing crowd running a local LLM on a MacBook, this is a genuinely exciting week.

• The gain: up to ~90% faster token generation for Gemma 4 on Apple Silicon
• How: multi-token prediction plus automatic performance tuning
• Also in v0.32.0: tighter Gemma 4 model loading and refreshed MLX and llama.cpp engines
• The lineup: Gemma 4 ships in E2B (phones), E4B (edge), 12B Unified (multimodal with audio), 26B MoE, and 31B Dense

Why Does Multi-Token Prediction Speed Things Up So Much?

Traditional decoding produces one token at a time — a strict, sequential bottleneck. Multi-token prediction lets the model propose several tokens per step and verify them together, so more of your Mac's memory bandwidth and Neural Engine headroom gets used on each pass. Pair that with automatic tuning that adapts to your specific chip, and you get a large speedup without changing a single line of your prompt. The refreshed MLX and llama.cpp back ends do the heavy lifting under the hood.

Which Gemma 4 Size Should You Run Locally?

That depends on your hardware and your goal. The E2B and E4B variants are built for phones and edge devices and sip resources. The new 12B Unified model is the sweet spot for a modern laptop — multimodal, now with audio, and practical to keep resident in memory. The 26B Mixture-of-Experts model activates only a fraction of its parameters per token, giving consumer GPUs strong quality at manageable speed, while the 31B Dense model targets workstations.

This is the same local-first momentum we track across our AI coverage, and it dovetails with hardware trends too — faster on-device inference makes compact machines far more capable, a theme we explored in PyTorch 2.13's FlexAttention on Apple Silicon. If you are shopping for a machine to run models like these, our companion guide on the best mini PCs for local LLMs walks through the memory and bandwidth that matter most. Open weights plus a faster runtime is a wonderful combination — and it keeps getting better.

Sources: Ollama Blog — July 11, 2026; PromptQuorum — Ollama July 2026 — July 2026; AurigaIT — Gemma 4 Guide — 2026.