Advisor / Hardware / Mac Studio M5 Ultra 512GB (ships late Oct 2026)

Best local LLMs for Mac Studio M5 Ultra 512GB (ships late Oct 2026)

Mac Studio — M5 Ultra (80-core GPU) 512GB. Memory-fit math with KV cache, speeds labelled by how they were obtained, and commands you can paste.

357.9 GBusable for model + KV cache
1200 GB/smemory bandwidth
0measured data points on file
Sponsor slot. No sponsor yet and no link.

Best for Coding · 16k context

ModelQuantFitMemorySpeed (generation)
GLM-4.5-AirQ8_0fits fully120.5 GB~47 tok/s estimated
Gemma 4 31BQ8_0fits fully35.3 GB~23 tok/s estimated
Qwen3.6 27BQ8_0fits fully29.8 GB~26 tok/s estimated
gpt-oss 120BQ8_0fits fully64.0 GB~215 tok/s estimated
Qwen3-Coder 30B-A3BQ8_0fits fully34.1 GB~169 tok/s estimated
llama.cpp · top pick
llama-server -hf unsloth/GLM-4.5-Air-GGUF:Q8_0 -c 16384 -fa on --jinja -ngl 99
ollama · top pick
OLLAMA_CONTEXT_LENGTH=16384 ollama run hf.co/unsloth/GLM-4.5-Air-GGUF:Q8_0

Best for Chat / general · 16k context

ModelQuantFitMemorySpeed (generation)
GLM-4.5-AirQ8_0fits fully120.5 GB~47 tok/s estimated
Gemma 4 31BQ8_0fits fully35.3 GB~23 tok/s estimated
Qwen3.6 27BQ8_0fits fully29.8 GB~26 tok/s estimated
gpt-oss 120BQ8_0fits fully64.0 GB~215 tok/s estimated
Mistral Small 3.2 24BQ8_0fits fully27.7 GB~30 tok/s estimated
llama.cpp · top pick
llama-server -hf unsloth/GLM-4.5-Air-GGUF:Q8_0 -c 16384 -fa on --jinja -ngl 99
ollama · top pick
OLLAMA_CONTEXT_LENGTH=16384 ollama run hf.co/unsloth/GLM-4.5-Air-GGUF:Q8_0

Best for Vision · 16k context

ModelQuantFitMemorySpeed (generation)
Gemma 4 31BQ8_0fits fully35.3 GB~23 tok/s estimated
Qwen3.6 27BQ8_0fits fully29.8 GB~26 tok/s estimated
Mistral Small 3.2 24BQ8_0fits fully27.7 GB~30 tok/s estimated
Gemma 4 12BQ8_0fits fully13.7 GB~58 tok/s estimated
Qwen3.6 35B-A3BQ8_0fits fully37.3 GB~193 tok/s estimated
llama.cpp · top pick
llama-server -hf unsloth/gemma-4-31B-it-GGUF:Q8_0 -c 16384 -fa on --jinja -ngl 99
ollama · top pick
OLLAMA_CONTEXT_LENGTH=16384 ollama run hf.co/unsloth/gemma-4-31B-it-GGUF:Q8_0

All 13 models that fit, fastest first · 16k context

Click a column header to sort. 0 of 13 speeds here are measured or scaled from measured; the rest are bandwidth estimates. Filter by use case, speed source, fit, license and more in the interactive advisor →

ModelParamsQuantFitMemoryTok/sSource
gpt-oss 20B20.91B MoEQ8_0fits fully12.5 GB~285estimated
gpt-oss 120B116.83B MoEQ8_0fits fully64.0 GB~215estimated
Qwen3.6 35B-A3B35.95B MoEQ8_0fits fully37.3 GB~193estimated
Qwen3-Coder 30B-A3B30.53B MoEQ8_0fits fully34.1 GB~169estimated
Qwen3.5 4B4.66BQ8_0fits fully5.1 GB~169estimated
Gemma 4 26B-A4B25.81B MoEQ8_0fits fully27.9 GB~148estimated
Qwen3-VL 8B8.77BQ8_0fits fully11.1 GB~87estimated
Qwen3.5 9B9.65BQ8_0fits fully10.1 GB~79estimated
Gemma 4 12B11.96BQ8_0fits fully13.7 GB~58estimated
GLM-4.5-Air110.47B MoEQ8_0fits fully120.5 GB~47estimated
Mistral Small 3.2 24B24.01BQ8_0fits fully27.7 GB~30estimated
Qwen3.6 27B27.78BQ8_0fits fully29.8 GB~26estimated
Gemma 4 31B31.27BQ8_0fits fully35.3 GB~23estimated

Long context on this hardware · up to 1M tokens

Largest context whose KV cache still fits next to the weights (same placement as the 16k result, KV types f16 / q8_0 / q4_0). Models are capped at their verified maximum; extended means beyond the native maximum using YaRN, only listed where the model card documents it. Long context needs flash attention (-fa on, already in the commands) and a lot of memory; a quantized KV cache saves memory but can cost accuracy. Speed at depth is an estimate (penalty fitted to measured depth runs up to 128k).

Try it in the advisor: 128k · 256k · 512k · 1M

ModelNative maxMax ctx f16Max ctx q8_0Max ctx q4_0Est. tok/s at max f16 ctx
Qwen3.6 35B-A3B Q8_0256k1M extended1M extended1M extended~33 estimated
Qwen3-Coder 30B-A3B Q8_0256k1M extended1M extended1M extended~8 estimated
Qwen3.5 4B Q8_0256k1M extended1M extended1M extended~27 estimated
Qwen3-VL 8B Q8_0256k1M extended1M extended1M extended~7 estimated
Qwen3.5 9B Q8_0256k1M extended1M extended1M extended~23 estimated
Qwen3.6 27B Q8_0256k1M extended1M extended1M extended~10 estimated
Gemma 4 26B-A4B Q8_0256k256k256k256k~74 estimated
Gemma 4 12B Q8_0256k256k256k256k~46 estimated
Gemma 4 31B Q8_0256k256k256k256k~15 estimated
gpt-oss 20B Q8_0128k128k128k128k~135 estimated
gpt-oss 120B Q8_0128k128k128k128k~95 estimated
GLM-4.5-Air Q8_0128k128k128k128k~19 estimated
Mistral Small 3.2 24B Q8_0128k128k128k128k~19 estimated
Affiliate box: “cheapest upgrade” for Mac Studio M5 Ultra 512GB (ships late Oct 2026). No affiliate links yet.

Open the interactive advisor with this hardware →

Speeds marked measured are medians of real runs (short-context generation) from these sources; scaled from measured means a measurement of a same-size or same-model-different-quant run, scaled by bytes read; estimated comes from memory bandwidth (spec source: https://support.apple.com/en-us/128107) × an efficiency factor of 0.63 calibrated to measured runs on this hardware class. No measured data on file for this hardware yet · submit your benchmark. Memory size on this page is an example configuration; check what your machine has. Model data last verified 2026-10-14. Ranking = fits and fast enough first, then larger; not a quality benchmark.

Page generated 2026-10-16 by tools/generate_pages.js