Best LLM for 96 GB unified memory · 2026

Best LLM for a 96 GB Mac: what fits, and what to run

Updated 8 Oct 2026 1 new this month

macOS lets the GPU wire down about 72.0 GB of a 96 GB Mac's unified memory; the rest stays with the system and your apps. At Q4_K_M with an 8K context, 66 of 80 models in the catalogue fit — 60 with headroom, 6 tightly. The pick for most people is Qwen3.8 27B: Best capability per gigabyte: a current-generation dense 27B with vision and 262K context. Speeds are estimated for an M2 Max · 96 GB; the fit verdicts are the same for every machine in the class.

Machines in this class M2 Max · 96 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 Qwen3.8 27B
Alibaba · Apache 2.0
Best capability per gigabyte: a current-generation dense 27B with vision and 262K context. Q4_K_M · 16.7 GB 14 Runs great
Best for coding Qwen3.8 27B
Alibaba · Apache 2.0
Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB. Q4_K_M · 16.7 GB 14 Runs great
Best reasoning Gemma 4 31B
Google · Apache 2.0
89% AIME — the strongest maths of any model under 32B. Q4_K_M · 20.2 GB 13 Runs great
Best for writing Llama 3.3 70B Instruct
Meta · Llama 3.3 Community
The creative-writing favourite: consistent voice, takes direction, handles fiction framing. Q4_K_M · 42.8 GB 5.6 Runs great
Best vision Qwen3.8 27B
Alibaba · Apache 2.0
Image and video in, OSWorld 84 — the strongest local vision-language model at 24 GB. Q4_K_M · 16.7 GB 14 Runs great
Best for agents Qwen3.8 27B
Alibaba · Apache 2.0
Computer use and tool calling are what 3.8 was trained for (OSWorld 84). Q4_K_M · 16.7 GB 14 Runs great
Best translation Qwen3.8 27B
Alibaba · Apache 2.0
119 languages, and Qwen translations read as idiomatic rather than literal. Q4_K_M · 16.7 GB 14 Runs great
Fastest good model Qwen3.6 35B-A3B
Alibaba · Apache 2.0
3.3B active per token: the fastest model that still competes with dense 27Bs. Q4_K_M · 20.9 GB 56 Runs great
Best long context Qwen3.8 27B
Alibaba · Apache 2.0
262K native, and only 16 of 64 blocks keep a KV cache — long context is cheap here. Q4_K_M · 24.2 GB
at 128K context
14 Runs great

The long-context row is judged at 128K: Qwen3.8 27B holds that window on a 96 GB Mac at Q4_K_M with 8.00 GB of KV cache — that is why it can differ from the best overall pick.

Everything that fits a 96 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.
Llama 3.3 70B Instruct 70.6B Q4_K_M 39.7 GB 2.50 GB 42.8 GB 29.2 GB 101K 5.6 Runs great
DeepSeek-R1-Distill-Llama 70B 70.6B Q4_K_M 39.7 GB 2.50 GB 42.8 GB 29.2 GB 101K 5.6 Runs great
Qwen3.6 35B-A3B 35.9B MoE Q4_K_M 20.2 GB 0.16 GB 20.9 GB 51.1 GB 256K 56 Runs great
Ornith 1.5 35B-A3B 35.9B MoE Q4_K_M 20.2 GB 0.16 GB 20.9 GB 51.1 GB 256K 56 Runs great
LLM-jp 4 33B Thinking 33.2B Q4_K_M 18.7 GB 2.00 GB 21.3 GB 50.7 GB 64K 12 Runs great
Qwen3 32B 32.8B Q4_K_M 18.4 GB 2.00 GB 21.0 GB 51.0 GB 128K 12 Runs great
DeepSeek-R1-Distill-Qwen 32B 32.8B Q4_K_M 18.4 GB 2.00 GB 21.0 GB 51.0 GB 128K 12 Runs great
Qwen2.5-Coder 32B 32.8B Q4_K_M 18.4 GB 2.00 GB 21.0 GB 51.0 GB 128K 12 Runs great
Olmo 3.1 32B Instruct 32.2B Q4_K_M 18.1 GB 1.25 GB 20.0 GB 52.0 GB 64K 12 Runs great
Nemotron 3.5 Lightning 30B-A3B 31.6B MoE Q4_K_M 17.8 GB 0.05 GB 18.4 GB 53.6 GB 256K 58 Runs great
Gemma 4 31B 31.3B Q4_K_M 17.6 GB 2.03 GB 20.2 GB 51.8 GB 256K 13 Runs great
GLM-4.7-Flash 30B-A3B 31.2B MoE Q4_K_M 17.5 GB 0.41 GB 18.6 GB 53.4 GB 198K 62 Runs great
Qwen3 30B-A3B 30.5B MoE Q4_K_M 17.1 GB 0.75 GB 18.5 GB 53.5 GB 128K 56 Runs great
Qwen3 Coder 30B-A3B 30.5B MoE Q4_K_M 17.1 GB 0.75 GB 18.5 GB 53.5 GB 256K 56 Runs great
Muse Glimmer 30B 29.8B Q4_K_M 16.8 GB 0.18 GB 17.5 GB 54.5 GB 128K 13 Runs great
Granite 4.1 30B 28.9B Q4_K_M 16.3 GB 2.00 GB 18.9 GB 53.1 GB 128K 14 Runs great
Qwen3.8 27B 27.8B Q4_K_M 15.6 GB 0.50 GB 16.7 GB 55.3 GB 256K 14 Runs great
Qwen3.6 27B 27.8B Q4_K_M 15.6 GB 0.50 GB 16.7 GB 55.3 GB 256K 14 Runs great
Gemma 3 27B 27.4B Q4_K_M 15.4 GB 1.03 GB 17.0 GB 55.0 GB 128K 15 Runs great
Gemma 4 26B-A4B 26.5B MoE Q4_K_M 14.9 GB 0.51 GB 16.0 GB 56.0 GB 256K 49 Runs great
Devstral Small 2 24B 24B Q4_K_M 13.5 GB 1.25 GB 15.3 GB 56.7 GB 370K 17 Runs great
Mistral Small 3.2 24B 23.6B Q4_K_M 13.3 GB 1.25 GB 15.1 GB 56.9 GB 128K 17 Runs great
gpt-oss 20B 20.9B MoE MXFP4 10.8 GB 0.19 GB 11.6 GB 60.4 GB 128K 56 Runs great
Qwen3 14B 14.8B Q4_K_M 8.3 GB 1.25 GB 10.2 GB 61.8 GB 128K 27 Runs great
DeepSeek-R1-Distill-Qwen 14B 14.8B Q4_K_M 8.3 GB 1.50 GB 10.4 GB 61.6 GB 128K 27 Runs great
Qwen2.5-Coder 14B 14.8B Q4_K_M 8.3 GB 1.50 GB 10.4 GB 61.6 GB 128K 27 Runs great
Phi-4 14B 14.7B Q4_K_M 8.3 GB 1.56 GB 10.4 GB 61.6 GB 16K 27 Runs great
Ministral 3 14B 13.9B Q4_K_M 7.8 GB 1.25 GB 9.7 GB 62.3 GB 256K 29 Runs great
Gemma 3 12B 12.2B Q4_K_M 6.9 GB 0.81 GB 8.3 GB 63.7 GB 128K 33 Runs great
Mistral NeMo 12B 12.2B Q4_K_M 6.9 GB 1.25 GB 8.7 GB 63.3 GB 128K 33 Runs great
Gemma 4 12B 12B Q4_K_M 6.7 GB 0.81 GB 8.2 GB 63.8 GB 256K 33 Runs great
Qwen3.5 9B 9.65B Q4_K_M 5.4 GB 0.25 GB 6.3 GB 65.7 GB 256K 41 Runs great
Ornith 1.5 9B 9.41B Q4_K_M 5.3 GB 0.25 GB 6.1 GB 65.9 GB 256K 42 Runs great
Ministral 3 8B 8.92B Q4_K_M 5.0 GB 1.06 GB 6.7 GB 65.3 GB 256K 45 Runs great
Granite 4.1 8B 8.79B Q4_K_M 4.9 GB 1.25 GB 6.8 GB 65.2 GB 128K 45 Runs great
LFM2.5 8B-A1B 8.47B MoE Q4_K_M 4.8 GB 0.09 GB 5.5 GB 66.5 GB 125K 124 Runs great
Fara 7B 8.29B Q4_K_M 4.7 GB 0.44 GB 5.7 GB 66.3 GB 125K 48 Runs great
Qwen3 8B 8.19B Q4_K_M 4.6 GB 1.13 GB 6.3 GB 65.7 GB 128K 49 Runs great
Llama 3.1 8B Instruct 8.03B Q4_K_M 4.5 GB 1.00 GB 6.1 GB 65.9 GB 128K 50 Runs great
Gemma 4 E4B 8.0B Q4_K_M 4.5 GB 0.14 GB 5.2 GB 66.8 GB 128K 50 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 66.9 GB 128K 143 Runs great
DeepSeek-R1-Distill-Qwen 7B 7.62B Q4_K_M 4.3 GB 0.44 GB 5.3 GB 66.7 GB 128K 52 Runs great
Qwen2.5-Coder 7B 7.62B Q4_K_M 4.3 GB 0.44 GB 5.3 GB 66.7 GB 128K 52 Runs great
Olmo 3 7B Instruct 7.3B Q4_K_M 4.1 GB 2.50 GB 7.2 GB 64.8 GB 64K 54 Runs great
Mistral 7B Instruct v0.3 7.25B Q4_K_M 4.1 GB 1.00 GB 5.7 GB 66.3 GB 32K 55 Runs great
Gemma 4 E2B 5.1B Q4_K_M 2.9 GB 0.07 GB 3.5 GB 68.5 GB 128K 78 Runs great
Qwen3.5 4B 4.66B Q4_K_M 2.6 GB 0.25 GB 3.5 GB 68.5 GB 256K 85 Runs great
Gemma 3 4B 4.3B Q4_K_M 2.4 GB 0.30 GB 3.3 GB 68.7 GB 128K 92 Runs great
Qwen3 4B 4.02B Q4_K_M 2.3 GB 1.13 GB 4.0 GB 68.0 GB 32K 99 Runs great
Ministral 3 3B 3.85B Q4_K_M 2.2 GB 0.81 GB 3.6 GB 68.4 GB 256K 103 Runs great
Phi-4-mini 3.8B 3.84B Q4_K_M 2.2 GB 1.00 GB 3.8 GB 68.2 GB 128K 104 Runs great
Granite 4.1 3B 3.4B Q4_K_M 1.9 GB 0.63 GB 3.1 GB 68.9 GB 128K 117 Runs great
Llama 3.2 3B Instruct 3.21B Q4_K_M 1.8 GB 0.88 GB 3.3 GB 68.7 GB 128K 124 Runs great
LFM2.5 2.6B 2.7B Q4_K_M 1.5 GB 0.13 GB 2.2 GB 69.8 GB 128K 147 Runs great
Qwen3.5 2B 2.27B Q4_K_M 1.3 GB 0.09 GB 2.0 GB 70.0 GB 256K 175 Runs great
Qwen3 1.7B 1.72B Q4_K_M 1.0 GB 0.88 GB 2.4 GB 69.6 GB 32K 231 Runs great
Llama 3.2 1B Instruct 1.24B Q4_K_M 0.7 GB 0.25 GB 1.5 GB 70.5 GB 128K 321 Runs great
Gemma 3 1B 1.0B Q4_K_M 0.6 GB 0.04 GB 1.2 GB 70.8 GB 32K 398 Runs great
Qwen3.5 0.8B 0.87B Q4_K_M 0.5 GB 0.09 GB 1.2 GB 70.8 GB 256K 457 Runs great
Qwen3 0.6B 0.6B Q4_K_M 0.3 GB 0.88 GB 1.8 GB 70.2 GB 32K 663 Runs great
Qwen3.5 122B-A10B 125B MoE Q4_K_M 70.3 GB 0.19 GB 71.1 GB 0.9 GB 47K 19 Tight
Nemotron 3 Super 120B-A12B 124B MoE Q4_K_M 69.7 GB 0.06 GB 70.4 GB 1.6 GB 214K 15 Tight
Ling 3.0 Flash 124B-A5B 124B MoE Q4_K_M 69.7 GB 0.06 GB 70.4 GB 1.6 GB 218K 36 Tight
Mistral Small 4 119B-A6B 119B MoE Q4_K_M 66.9 GB 0.18 GB 67.7 GB 4.3 GB 204K 29 Tight
gpt-oss 120B 117B MoE MXFP4 60.5 GB 0.29 GB 61.4 GB 10.6 GB 128K 39 Tight
Llama 4 Scout 109B-A17B 109B MoE Q4_K_M 61.3 GB 1.50 GB 63.4 GB 8.6 GB 191K 11 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
Qwen3.8-Flash-Next 180B-A6B 180B MoE 102.0 GB +30.0 GB 192 GB Mac
Qwen3 235B-A22B 235B MoE 134.2 GB +62.2 GB 256 GB Mac
DeepSeek V4 Flash 284B-A13B 284B MoE 160.7 GB +88.7 GB 256 GB Mac
GLM-5.3-Flash 320B-A18B 321B MoE 181.2 GB +109.2 GB 512 GB Mac
Ornith 1.5 397B-A17B 397B MoE 224.1 GB +152.1 GB 512 GB Mac
Qwen3.5 397B-A17B 403B MoE 227.4 GB +155.4 GB 512 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 (M2 Max 400 GB/s), and tokens per second scales with it.


← 64 GB Mac All memory classes 128 GB Mac → Adjust context, quant or cache →

Every total is quantised weights + f16 KV cache at 8K + 0.6 GB runtime overhead, against 72.0 GB usable. Tokens per second are estimated from memory bandwidth and never measured. The method.