Best LLM for 16 GB unified memory · 2026

Best LLM for a 16 GB Mac: what fits, and what to run

Updated 8 Oct 2026 1 new this month

macOS lets the GPU wire down about 10.7 GB of a 16 GB Mac's unified memory; the rest stays with the system and your apps. At Q4_K_M with an 8K context, 37 of 80 models in the catalogue fit — 32 with headroom, 5 tightly. The pick for most people is Gemma 4 12B: The 12 GB generalist: text, image and audio in, 140+ languages. Speeds are estimated for an M4 · 16 GB; the fit verdicts are the same for every machine in the class.

Machines in this class M4 · 16 GBM1 · 16 GB

Which one should I run?

One pick per job, from the editorial shortlist, checked to fit at Q4_K_M and 8K. Each label links the full ranking for that job. How picks are chosen.

RecommendationModelWhyQuant · totalTok/s est.
Best overall Gemma 4 12B
Google · Apache 2.0
The 12 GB generalist: text, image and audio in, 140+ languages. Q4_K_M · 8.2 GB 9.9 Runs great
Best for coding Qwen3.5 9B
Alibaba · Apache 2.0
The best coding model for 8–12 GB cards. Q4_K_M · 6.3 GB 12 Runs great
Best reasoning Qwen3.5 9B
Alibaba · Apache 2.0
Thinking mode on a 12 GB card — ahead of the 2025 R1 distills. Q4_K_M · 6.3 GB 12 Runs great
Best for writing Gemma 4 12B
Google · Apache 2.0
The 12 GB writing pick; 140+ languages. Q4_K_M · 8.2 GB 9.9 Runs great
Best vision Gemma 4 12B
Google · Apache 2.0
Image and audio understanding on 12 GB. Q4_K_M · 8.2 GB 9.9 Runs great
Best for agents Granite 4.1 8B
IBM · Apache 2.0
Enterprise-grade tool calling in a dense 8B, no thinking overhead. Q4_K_M · 6.8 GB 14 Runs great
Best translation Gemma 4 12B
Google · Apache 2.0
Broad language coverage on a 12 GB card. Q4_K_M · 8.2 GB 9.9 Runs great
Fastest good model LFM2.5 8B-A1B
Liquid AI · LFM Open License v1.0
1.5B active; designed for laptops without a GPU. Q4_K_M · 5.5 GB 37 Runs great
Best long context Qwen3.5 4B
Alibaba · Apache 2.0
262K on 8 GB. Q4_K_M · 7.2 GB
at 128K context
26 Runs great

The long-context row is judged at 128K: Qwen3.5 4B holds that window on a 16 GB Mac at Q4_K_M with 4.00 GB of KV cache — that is why it can differ from the best overall pick.

Everything that fits a 16 GB Mac

Largest first. "Max context" is the longest window the Mac holds at that quantisation with an f16 cache; a q8_0 cache roughly doubles it.

ModelParamsQuantWeights+KV 8KTotalHeadroomMax contextTok/s est.
Gemma 3 12B 12.2B Q4_K_M 6.9 GB 0.81 GB 8.3 GB 2.4 GB 46K 9.8 Runs great
Mistral NeMo 12B 12.2B Q4_K_M 6.9 GB 1.25 GB 8.7 GB 2.0 GB 20K 9.8 Runs great
Gemma 4 12B 12B Q4_K_M 6.7 GB 0.81 GB 8.2 GB 2.5 GB 48K 9.9 Runs great
Qwen3.5 9B 9.65B Q4_K_M 5.4 GB 0.25 GB 6.3 GB 4.4 GB 149K 12 Runs great
Ornith 1.5 9B 9.41B Q4_K_M 5.3 GB 0.25 GB 6.1 GB 4.6 GB 153K 13 Runs great
Ministral 3 8B 8.92B Q4_K_M 5.0 GB 1.06 GB 6.7 GB 4.0 GB 38K 13 Runs great
Granite 4.1 8B 8.79B Q4_K_M 4.9 GB 1.25 GB 6.8 GB 3.9 GB 33K 14 Runs great
LFM2.5 8B-A1B 8.47B MoE Q4_K_M 4.8 GB 0.09 GB 5.5 GB 5.2 GB 125K 37 Runs great
Fara 7B 8.29B Q4_K_M 4.7 GB 0.44 GB 5.7 GB 5.0 GB 99K 14 Runs great
Qwen3 8B 8.19B Q4_K_M 4.6 GB 1.13 GB 6.3 GB 4.4 GB 39K 15 Runs great
Llama 3.1 8B Instruct 8.03B Q4_K_M 4.5 GB 1.00 GB 6.1 GB 4.6 GB 44K 15 Runs great
Gemma 4 E4B 8.0B Q4_K_M 4.5 GB 0.14 GB 5.2 GB 5.5 GB 128K 15 Runs great
Ling 3.0 Tiny 7.9B-A1.3B 7.9B MoE Q4_K_M 4.4 GB 0.05 GB 5.1 GB 5.6 GB 128K 43 Runs great
DeepSeek-R1-Distill-Qwen 7B 7.62B Q4_K_M 4.3 GB 0.44 GB 5.3 GB 5.4 GB 106K 16 Runs great
Qwen2.5-Coder 7B 7.62B Q4_K_M 4.3 GB 0.44 GB 5.3 GB 5.4 GB 106K 16 Runs great
Olmo 3 7B Instruct 7.3B Q4_K_M 4.1 GB 2.50 GB 7.2 GB 3.5 GB 35K 16 Runs great
Mistral 7B Instruct v0.3 7.25B Q4_K_M 4.1 GB 1.00 GB 5.7 GB 5.0 GB 32K 16 Runs great
Gemma 4 E2B 5.1B Q4_K_M 2.9 GB 0.07 GB 3.5 GB 7.2 GB 128K 23 Runs great
Qwen3.5 4B 4.66B Q4_K_M 2.6 GB 0.25 GB 3.5 GB 7.2 GB 239K 26 Runs great
Gemma 3 4B 4.3B Q4_K_M 2.4 GB 0.30 GB 3.3 GB 7.4 GB 128K 28 Runs great
Qwen3 4B 4.02B Q4_K_M 2.3 GB 1.13 GB 4.0 GB 6.7 GB 32K 30 Runs great
Ministral 3 3B 3.85B Q4_K_M 2.2 GB 0.81 GB 3.6 GB 7.1 GB 78K 31 Runs great
Phi-4-mini 3.8B 3.84B Q4_K_M 2.2 GB 1.00 GB 3.8 GB 6.9 GB 63K 31 Runs great
Granite 4.1 3B 3.4B Q4_K_M 1.9 GB 0.63 GB 3.1 GB 7.6 GB 104K 35 Runs great
Llama 3.2 3B Instruct 3.21B Q4_K_M 1.8 GB 0.88 GB 3.3 GB 7.4 GB 75K 37 Runs great
LFM2.5 2.6B 2.7B Q4_K_M 1.5 GB 0.13 GB 2.2 GB 8.5 GB 128K 44 Runs great
Qwen3.5 2B 2.27B Q4_K_M 1.3 GB 0.09 GB 2.0 GB 8.7 GB 256K 53 Runs great
Qwen3 1.7B 1.72B Q4_K_M 1.0 GB 0.88 GB 2.4 GB 8.3 GB 32K 69 Runs great
Llama 3.2 1B Instruct 1.24B Q4_K_M 0.7 GB 0.25 GB 1.5 GB 9.2 GB 128K 96 Runs great
Gemma 3 1B 1.0B Q4_K_M 0.6 GB 0.04 GB 1.2 GB 9.5 GB 32K 119 Runs great
Qwen3.5 0.8B 0.87B Q4_K_M 0.5 GB 0.09 GB 1.2 GB 9.5 GB 256K 137 Runs great
Qwen3 0.6B 0.6B Q4_K_M 0.3 GB 0.88 GB 1.8 GB 8.9 GB 32K 199 Runs great
Qwen3 14B 14.8B Q4_K_M 8.3 GB 1.25 GB 10.2 GB 0.5 GB 11K 8.1 Tight
DeepSeek-R1-Distill-Qwen 14B 14.8B Q4_K_M 8.3 GB 1.50 GB 10.4 GB 0.3 GB 9K 8.1 Tight
Qwen2.5-Coder 14B 14.8B Q4_K_M 8.3 GB 1.50 GB 10.4 GB 0.3 GB 9K 8.1 Tight
Phi-4 14B 14.7B Q4_K_M 8.3 GB 1.56 GB 10.4 GB 0.3 GB 9K 8.1 Tight
Ministral 3 14B 13.9B Q4_K_M 7.8 GB 1.25 GB 9.7 GB 1.0 GB 14K 8.6 Tight

Just out of reach

The smallest models that do not fit. Unified memory has no slower pool to spill into, so there is no offload fallback: each needs a bigger machine, or a lower quantisation.

ModelParamsTotal @ Q4_K_MOver budgetNeeds
gpt-oss 20B 20.9B MoE 11.6 GB +0.9 GB 24 GB Mac
Mistral Small 3.2 24B 23.6B 15.1 GB +4.4 GB 24 GB Mac
Devstral Small 2 24B 24B 15.3 GB +4.6 GB 32 GB Mac
Gemma 4 26B-A4B 26.5B MoE 16.0 GB +5.3 GB 32 GB Mac
Qwen3.8 27B 27.8B 16.7 GB +6.0 GB 32 GB Mac
Qwen3.6 27B 27.8B 16.7 GB +6.0 GB 32 GB Mac

Running it on a Mac

The GPU's share can be raised. The cap macOS puts on wired GPU memory is a setting, not a hardware limit: sudo sysctl iogpu.wired_limit_mb=… lifts it until the next reboot. Leave several gigabytes for macOS itself — every number on this page assumes the default share, so it under-promises rather than over.

MLX or GGUF. On Apple Silicon, MLX (via mlx-lm or LM Studio) is usually faster than llama.cpp's GGUF on the same model, and the community converts most popular releases within days. GGUF through Ollama or llama.cpp has the widest coverage and usually supports a new architecture first. The memory arithmetic is the same either way.

Speed is bandwidth. Every machine in this class holds the same models; what separates them is memory bandwidth (M4 120 GB/s, M1 68 GB/s), and tokens per second scales with it.


← 8 GB Mac All memory classes 24 GB Mac → Adjust context, quant or cache →

Every total is quantised weights + f16 KV cache at 8K + 0.6 GB runtime overhead, against 10.7 GB usable. Tokens per second are estimated from memory bandwidth and never measured. The method.