Best local LLM for vision / image input on a Apple M3 Ultra (96GB) (2026)
All figures assume an f16 KV cache, a 0.6 GB display reserve on the GPU, and 64 GB of DDR5 system RAM for the offload tiers. Tune these in the calculator.
The verdict
Gemma 3 27B Q8_0
Fits on GPU at 16K · Vision / image input score 8/10 · ≈ 18 tok/s
Sliding-window attention keeps long context cheap, and the vision stack still beats most of 2026's newcomers.
llama-server -m gemma-3-27b-it-Q8_0.gguf -c 16384 --flash-attn -ngl 99
Worthy alternates
Gemma 4 26B A4B Q8_0
Fits on GPU · ≈ 79 tok/s · Vision / image input 8/10
Google's fast MoE with native audio in. Nearly all of its weight sits in routed experts, so expert offload runs it comfortably on 12 GB cards.
Gemma 4 12B Q8_0
Fits on GPU · ≈ 41 tok/s · Vision / image input 7/10
June 2026's 16GB-class headline: dense 12B with native vision and audio in one backbone. The new laptop ceiling.
Tune this for your exact RAM and settings in the calculator → · All models on the Apple M3 Ultra (96GB)
Frequently asked questions
What is the best local LLM for vision / image input on a Apple M3 Ultra (96GB)?
Gemma 3 27B at Q8_0 — it scores 8/10 for vision / image input and runs as "Fits on GPU" at 16K context on the Apple M3 Ultra (96GB).
How much context do I need for vision / image input?
We recommend 16K tokens for vision / image input (minimum 8K). These picks are computed at 16K.
How fast will it run on a Apple M3 Ultra (96GB)?
Roughly 18 tokens/sec for Gemma 3 27B — comfortable for interactive use.
Do I need more than 72 GB of VRAM for vision / image input?
No — the pick above needs 30.4 GB of VRAM at 16K.
What settings should I use?
Start with our command: llama-server -m gemma-3-27b-it-Q8_0.gguf -c 16384 --flash-attn -ngl 99 — then tune context and KV quant in the fit calculator.