The best local LLM for a Mac Studio M5 Ultra (96 GB) is KAT-Coder-V2.5 at 148 tok/s (estimated). With 96 GB of unified memory it runs 59 of the models in the LLMCheck index — from compact options up to 276B-class models. For everyday chat and coding, KAT-Coder-V2.5 is the sweet spot. Full ranking below.
Considering this Mac?
Mac Studio M5 Ultra (96GB) on Amazon → · not sure? the Mac Advisor picks a config →
Ranked by LLMCheck suitability (capability balanced against speed on the M5 Ultra). Click a model for its full benchmark and setup. Every speed carries its provenance mark: most are index estimates from the memory-bandwidth model, and a filled square is a published measurement — submit a real run →
| # | Model | Size | License | Speed | Capability |
|---|---|---|---|---|---|
| 1 | KAT-Coder-V2.5 | 35B | Apache 2.0 | estimated148 tok/s | 39/50 |
| 2 | Qwen 3.6-35B-A3B | 35B | Apache 2.0 | estimated155 tok/s | 38/50 |
| 3 | Qwen3.8-27B | 27.8B | Apache 2.0 | sourced55 tok/s | 46/50 |
| 4 | Inkling-Small | 276B | Apache 2.0 | estimated49 tok/s | 46/50 |
| 5 | Laguna S 2.1 | 118B | OpenMDW | estimated68 tok/s | 43/50 |
| 6 | Nemotron 3.5 Lightning | 30B | OpenMDW | estimated136 tok/s | 36/50 |
| 7 | Qwen 3.6-27B | 27B | Apache 2.0 | estimated57 tok/s | 44/50 |
| 8 | Qwen3-Coder-Next | 80B | Apache 2.0 | estimated130 tok/s | 35/50 |
| 9 | Gemma 4 26B-A4B | 26B | Apache 2.0 | estimated139 tok/s | 35/50 |
| 10 | Muse Glimmer 30B | 30B | Apache 2.0 | estimated52 tok/s | 42/50 |
| 11 | Laguna XS 2.1 | 33B | OpenMDW | estimated148 tok/s | 33/50 |
| 12 | Qwen 3.5 122B-A10B | 122B | Apache 2.0 | sourced80 tok/s | 38/50 |
Showing the top 12 of 59 models that fit in 96 GB. See the full leaderboard or all benchmarks.
The fastest way to get started is Ollama. Install it, then find the top pick for your Mac in its library and copy the exact tag it lists:
Find KAT-Coder-V2.5 in the Ollama library → · model card on Hugging Face →
Prefer a GUI? LM Studio gives you a one-click download and chat window. For step-by-step help see our Ollama install guide, or open the KAT-Coder-V2.5 on M5 Ultra benchmark page for exact settings.
The Mac Studio M5 Ultra (96 GB) comfortably runs 59 of the models in the LLMCheck index, led by KAT-Coder-V2.5. Grab one and start running LLMs offline today:
As an Amazon Associate, LLMCheck earns from qualifying purchases. Affiliate links cost you nothing extra and never influence our rankings.
KAT-Coder-V2.5 (35B, Apache 2.0) is the best all-round pick at 148 tok/s (estimated) on the M5 Ultra. If you want maximum speed, Maple Preview 20B-A1B hits 503 tok/s (estimated); for maximum capability, Qwen3.8-27B still fits in 96 GB.
About 59 of the 91 models in the LLMCheck leaderboard fit in 96 GB of unified memory, from compact models up to Inkling-Small (276B).
Yes. A 70B model in Q4 quantization needs roughly 40–44 GB of memory, which fits in 96 GB with headroom for context.
96 GB is plenty for local AI — you can run capable 30B–70B-class models. Because Apple Silicon uses unified memory, that figure is both your system RAM and your VRAM.