Best local LLM for summarization on a Apple M3 Pro (36GB) (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 Q6_K
Fits on GPU at 32K · Summarization score 8/10 · ≈ 4 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-Q6_K.gguf -c 32768 --flash-attn -ngl 99
Worthy alternates
Qwen3.5 27B Q4_K_M
Fits on GPU · ≈ 6 tok/s · Summarization 8/10
The dense 24GB workhorse. If you want one model on a 3090 and no surprises, it's this.
Qwen3.5 35B A3B Q4_K_M
Fits on GPU · ≈ 27 tok/s · Summarization 8/10
The meta pick, full stop. Near-dense-30B quality at 3B-active speed, and expert offload puts it on 8 GB cards.
Tune this for your exact RAM and settings in the calculator → · All models on the Apple M3 Pro (36GB)
Frequently asked questions
What is the best local LLM for summarization on a Apple M3 Pro (36GB)?
Gemma 3 27B at Q6_K — it scores 8/10 for summarization and runs as "Fits on GPU" at 32K context on the Apple M3 Pro (36GB).
How much context do I need for summarization?
We recommend 48K tokens for summarization (minimum 24K). These picks are computed at 32K.
How fast will it run on a Apple M3 Pro (36GB)?
Roughly 4 tokens/sec for Gemma 3 27B — usable for interactive use.
Do I need more than 27 GB of VRAM for summarization?
No — the pick above needs 25.8 GB of VRAM at 32K.
What settings should I use?
Start with our command: llama-server -m gemma-3-27b-it-Q6_K.gguf -c 32768 --flash-attn -ngl 99 — then tune context and KV quant in the fit calculator.