The M3 Ultra Mac Studio runs local models at 819 GB/s of memory bandwidth with 96 to 512 GB of unified memory. On Apple Silicon that memory is shared with the GPU, so the whole pool is available for weights: at 512 GB you can hold roughly a 696B dense model at Q4. Two Max dies fused together: the widest memory bus and the highest capacity Apple sells. This is the tier that runs frontier-size open weights locally.
Apple no longer sells this configuration new. It stays fully evaluated here because the used market is where most of its local AI value now sits.
Memory bandwidth is faster than 90% of the Apple Silicon chips shipped in a Mac, against a 1228 GB/s peak.
Its memory ceiling is above 90% of them, against a 512 GB peak.
Every option Apple sells with this chip. The model list below recomputes against the one you pick.
GPU cores
Unified memory
Apple couples memory to the core count on this chip, so the options change with the bin above.
Unified memory is the ceiling and it is soldered, so this is the decision you cannot revisit.
6,384 of 6,563 models fit, and 6,271 of them run with headroom rather than as a squeeze.
6,493 of 6,563 models fit, and 6,424 of them run with headroom rather than as a squeeze.
6,549 of 6,563 models fit, and 6,490 of them run with headroom rather than as a squeeze.
Every model in the database against this exact configuration, at 819 GB/s. Ratings and speeds are the same numbers the model pages show.
Showing 6563 of 6563 models
General · deepreinforce-ai · 2026-06-21
General · radixark · 2026-07-27
General · bahushruth · 2026-06-11
General · openbmb · 2026-05-21
General · farbodtavakkoli · 2026-06-17
General · internscience · 2026-07-13
General · Liquid AI · 2026-05-28
General · flywheel-ai · 2026-06-21
General · weiboai · 2026-06-12
General · Liquid AI · 2026-07-28
General · Alibaba · 2026-06-22
General · poolside · 2026-06-20
General · Liquid AI · 2026-06-24
General · nanbeige · 2026-07-21
General · ma7ee7 · 2026-07-30
General · apodex · 2026-06-07
General · internscience · 2026-06-22
Coding · coherelabs · 2026-06-05
General · goekdeniz-guelmez · 2026-07-31
General · ktruestory · 2026-05-28
General · petrouil · 2026-07-23
Multimodal · Alibaba · 2026-04-15
General · deepreinforce-ai · 2026-06-21
General · ibm-granite · 2026-04-06
Multimodal · ibm-granite · 2026-04-16
General · lukebailey181pub · 2026-04-21
General · trevorjs · 2026-04-03
General · empero-ai · 2026-06-19
General · poolside · 2026-04-23
General · huihui-ai · 2026-04-21
Multimodal · huihui-ai · 2026-04-18
General · goekdeniz-guelmez · 2026-07-31
Multimodal · Google · 2026-03-11
Multimodal · Alibaba · 2026-02-27
Multimodal · Google · 2026-03-02
General · ibm-granite · 2026-04-06
Multimodal · Google · 2026-03-02
Multimodal · google · 2026-05
Multimodal · Alibaba · 2026-02-28
Multimodal · Alibaba · 2026-02-28
Multimodal · Alibaba · 2026-02-24
Multimodal · Alibaba · 2026-02-28
Multimodal · Alibaba · 2026-02-28
General · yuxinlu1 · 2026-06-28
Multimodal · Alibaba · 2026-02-27
General · Liquid AI · 2026-03-31
General · ravichandranj · 2026-02-13
General · ibm-granite · 2026-04-16
General · nvidia · 2026-03-18
General · obliteratus · 2026-06-05
General · sarvamai · 2026-03-03
General · huihui-ai · 2026-07-11
General · obliteratus · 2026-04-15
General · nvidia · 2026-03-02
General · Liquid AI · 2026-02-24
Reasoning · jackrong · 2026-03-16
Reasoning · jackrong · 2026-03-07
General · cagrigungor · 2026-08-06
Multimodal · Alibaba · 2026-02-27
General · zai-org · 2026-01-19
Multimodal · Alibaba · 2026-02-26
General · nvidia · 2026-03-18
Multimodal · Liquid AI · 2026-01-05
General · Liquid AI · 2026-01-20
General · ai21labs · 2026-01-06
Reasoning · openonerec · 2026-06-09
General · Liquid AI · 2026-01-05
General · Liquid AI · 2026-01-04
General · Liquid AI · 2025-12-25
General · Liquid AI · 2026-01-05
Multimodal · davidau · 2026-02-02
Chat · bineric · 2026-01-12
General · inclusionai · 2026-02-09
General · Liquid AI · 2025-10-28
General · allenai · 2025-11-18
General · allenai · 2026-01-28
Chat · pearl-ai · 2026-02-26
Multimodal · Liquid AI · 2025-10-22
General · frontiersmind · 2026-08-03
General · alibaba-nlp · 2026-03-31
General · openai · 2025-08-04
Coding · Alibaba · 2025-07-31
Chat · Liquid AI · 2026-01-06
General · typhoon-ai · 2025-09-23
General · Alibaba · 2025-08-05
General · ibm-granite · 2025-09-16
General · ibm-granite · 2025-09-16
Multimodal · Liquid AI · 2025-08-12
General · allenai · 2025-09-12
General · openai · 2025-09-18
General · anton-hugging · 2026-02-06
General · z-lab · 2026-01-04
General · Alibaba · 2025-07-29
General · ibm-granite · 2025-09-16
General · openonerec · 2025-12-30
General · ibm-granite · 2025-09-16
General · bytedance · 2025-10-28
General · dreamfast · 2026-01-11
General · Liquid AI · 2025-10-07
General · z-lab · 2026-01-04
Multimodal · Liquid AI · 2025-08-12
General · Liquid AI · 2025-09-22
General · Liquid AI · 2025-09-30
General · Liquid AI · 2025-08-22
General · Liquid AI · 2025-09-03
General · Liquid AI · 2025-09-03
General · Liquid AI · 2025-08-25
General · Liquid AI · 2025-09-03
General · Liquid AI · 2025-09-03
General · NCAI · 2025-12-29
General · NCAI · 2025-12-29
Reasoning · DeepSeek · 2025-05-29
Chat · allenai · 2025-11-19
General · t-tech · 2025-12-22
General · hmellor · 2025-07-22
Chat · allenai · 2025-11-17
General · nvidia · 2025-08-12
General · inclusionai · 2025-11-25
General · ibm-granite · 2025-04-30
General · Liquid AI · 2025-07-10
Reasoning · Microsoft · 2025-04-29
Chat · xcuros · 2026-02-28
General · amd · 2025-05-17
General · stefanruseti · 2025-06-04
General · lgai-exaone · 2025-07-11
General · Liquid AI · 2025-07-10
General · menlo · 2025-06-25
General · Liquid AI · 2025-07-10
Multimodal · NCAI · 2025-12-29
Multimodal · Alibaba · 2025-01-26
General · huggingfacetb · 2025-07-08
General · Alibaba · 2025-09-23
General · nvidia · 2025-08-21
General · k-chirkunov · 2026-07-14
General · avitotech · 2025-10-20
General · huggingfacetb · 2025-06-19
General · Alibaba · 2025-09-23
General · nvidia · 2025-03-16
General · dream-org · 2025-04-03
General · openbmb · 2025-09-02
General · Alibaba · 2025-09-23
General · pfnet · 2025-02-05
General · bytedance-seed · 2025-04-09
Chat · baseten · 2025-09-12
General · sapientinc · 2026-05-17
General · ibm-granite · 2025-10-07
Chat · utter-project · 2026-01-26
General · fableforge-ai · 2026-07-05
General · viorikaai-org · 2026-07-05
General · bananamind · 2026-07-17
Both bins run the same 819 GB/s memory bus, so token generation is the same on either one. The extra cores show up in image and video work, not in tokens per second.
| Configuration | Memory bandwidth | Memory options | Models that fit |
|---|---|---|---|
| 28-core CPU, 60-core GPU | 819 GB/s | 96 GB | Identical |
| 32-core CPU, 80-core GPU | 819 GB/s | 96, 256, 512 GB | Identical |
Nobody has submitted a benchmark on the M3 Ultra yet, so every speed on this page is the formula estimate rather than a measured run. The estimate is bandwidth-driven and calibrated against chips that do have data, which makes it a good guide and not a promise.
ToolPiper contributes a result anonymously when you run the benchmark, and the leaderboard shows every chip that already has one.
What each step actually changes for local models, rather than which one is newer.
Apple has not announced any of this. The M7 Max and Ultra are the first parts Apple designed after cancelling a generation to reach them, so extrapolating from the M5 under-represents them. These rows assume LPDDR6, whose wider channels grow every bus by half, at its top speed bin by the time the Max and Ultra ship. The 1.5 TB Ultra ceiling is Bloomberg's reported design target, and whether that configuration ships depends on the memory market. Stacked memory or a new package fabric would land above these numbers; nobody outside Apple can price that yet.
Projected chip · expected 2028
2765 GB/s · 256 to 1536 GB unified memory
40-core CPU · 112 or 128-core GPU · 152 TOPS Neural Engine
Would hold about a 2092B model at Q4
On a PC the model has to fit in GPU VRAM, which is a separate pool from system RAM and usually the smaller of the two. Apple Silicon has one pool. The M3 Ultra's 819 GB/s bus is shared by CPU, GPU, and Neural Engine, so a 512 GB machine can hand almost all of that to a model with no copy across a bus.
Apple stopped selling this one, which is exactly why it is interesting. The Studio exists for this workload. It carries the widest memory buses and the highest capacities Apple sells, and it runs at full clocks indefinitely. A used M3 Ultra at 512 GB still gives you 819 GB/s and a hard 696B ceiling, and neither number degrades with age the way a battery does.
Yes, at 512 GB. A 70B model at Q4_K_M needs about 46 GB including an 8K context, and 512 GB of unified memory leaves about 450 GB for weights once macOS takes its share. At 96 GB it does not fit at any quantization worth running.
Memory is the only spec that changes what you can run at all. 96 GB holds about a 129B model at Q4; 512 GB holds about 696B. It is soldered, so this is a one-time decision, and it is the upgrade worth paying for before core count.
Token generation is bandwidth-bound, so M3 Ultra throughput scales with its 819 GB/s memory bus. Divide bandwidth by the size of the weights actually read per token to get the ceiling, then expect roughly half of that in practice. A 7B model at Q4 reads about 4 GB per token pass, so the M3 Ultra lands in the tens of tokens per second and a 70B model lands in the single digits.
Not for LLMs. Both bins run the same 819 GB/s memory bus and take the same memory options, and token generation is bound by bandwidth rather than GPU cores. The extra cores show up in image generation and video work, not in tokens per second.
For inference, the specs that matter do not age: 819 GB/s and up to 512 GB of unified memory are the same numbers today as they were in 2025. A used M3 Ultra at the top memory option usually beats a new base-tier machine at the same price on both. Check the battery and the display, not the silicon.
ToolPiper downloads, manages, and runs local models on Apple Silicon. Free, and nothing leaves the machine.