Best LLM for Mac Studio M5 Max with 36GB RAM (2026)

The 36GB Mac Studio M5 Max is the entry point into Apple's most powerful desktop for local AI in 2026. Announced on August 26 and shipping September 22, it starts at $2,499 with a binned 460 GB/s memory bus. That bandwidth is roughly 4x what a MacBook Air M4 delivers, and it can run 27B-class dense models and compact MoE architectures comfortably. This is the machine for anyone stepping up from a Mac Mini or MacBook Pro who wants desktop-class inference without paying for memory they do not need yet.

TL;DR: The 36GB M5 Max runs 27B dense and 35B MoE models at Q4. Qwen3.6 35B-A3B (~128 tok/s est.) tops the engine's ranking, followed by Gemma 4 26B-A4B (~84 tok/s est.) for multimodal work and Qwen3.8 27B (~23 tok/s est.), the best dense pick. Nemotron 3.5 Lightning 30B-A3B in its IQ4_XS build (~103 tok/s est.) and LFM2 24B-A2B (~141 tok/s est., the fastest pick) are the agent picks, and GLM-4.7-Flash (~90 tok/s est.) rounds out the six.

Estimated token generation on the Mac Studio M5 Max 36GB. ModelFit engine estimates.

The Mac Studio device page covers the full spec breakdown. This guide focuses on what actually runs well at 36GB, what to skip, and when to move up.

How 36GB Shapes Your Model Choices

macOS reserves memory for its kernel and services before any model can load. On a 36GB machine, your usable budget looks like this:

The rule of thumb for Q4_K_M quantization is roughly 0.6 GB per billion parameters. A 27B dense model needs about 16-18GB. A 35B MoE model stores all 35B of weights (~20-22GB at Q4) but activates only 3B per token. That is why MoE models decode so fast on bandwidth-bound hardware: memory moves only the active parameters each step.

The 36GB config uses a binned chip (32-core GPU instead of 40) on a 460 GB/s bus, not the full 614 GB/s found on higher tiers. Decode speed tracks memory bandwidth, so MoE and dense models alike run slower than on the 48GB tier: Qwen3.5 35B-A3B drops from ~166 to ~131 tok/s est. The trade-off is a meaningful savings at the checkout.

Best Models Ranked

Speeds are ModelFit estimates with a typical error of about 10%.

Model Details

1. Qwen3.6 35B-A3B is the engine's top pick for the 36GB tier. It stores 35B parameters but activates 3B per token, so it decodes at ~128 tok/s est. on the 460 GB/s bus. It loads ~22GB, which leaves about 3GB of the ~25GB budget for KV-cache: fine for interactive sessions at moderate context, while long-context work belongs on the 48GB M5 Max.

ollama run qwen3.6:35b-a3b

2. Gemma 4 26B-A4B is Google's most efficient large MoE. At ~16GB loaded and ~84 tok/s est., it is one of the lightest fits on this list and adds native image input. If you want one model that handles text, vision, and chat at high speed, this is the pick.

ollama run gemma4:26b

3. Qwen3.8 27B is the dense model champion at this tier. It needs ~16.5GB at Q4_K_M and runs at ~23 tok/s est. Its strength is long-context agentic coding, where the full 27B weights provide deep reasoning without MoE compression trade-offs.

ollama run qwen3.8:27b

4. Nemotron 3.5 Lightning 30B-A3B (IQ4_XS) is NVIDIA's hybrid Mamba-MoE: 30B stored, 3B active, and only 6 of its 52 layers use full attention, so the KV-cache stays a fraction of a dense 30B's and long contexts stay cheap. The Ollama :30b pull is a Q4_K_M of about 25GB that fills the ~25GB budget, so on 36GB use the bartowski IQ4_XS GGUF instead, 18.92GB on disk. It loads at ~18.9GB and runs at roughly 103 tok/s est., which leaves about 6GB for context. IQ4_XS is slightly lower precision than Q4_K_M, and Ollama has no tag for this build, so run it through llama.cpp:

llama-server -hf bartowski/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF:IQ4_XS

5. LFM2 24B-A2B Instruct is Liquid AI's agent MoE: 24B total parameters with about 2B active per token, built for function calling and multi-step agent pipelines (model card). It loads ~14GB, the lightest fit on this list, and at ~141 tok/s est. it is also the fastest.

ollama run lfm2:24b-a2b

6. GLM-4.7-Flash 30B-A3B is Z.ai's local-class MoE (MIT, 3B active). It scores 59.2 on SWE-bench Verified per its model card. The Ollama pull is a Q4_K_M of ~19GB and the engine estimates ~90 tok/s, which leaves about 6GB of the budget for context.

ollama run glm-4.7-flash

Earlier picks that now rank lower: Qwen3.5 35B-A3B Instruct (ollama run qwen3.5:35b-a3b, ~20GB, ~131 tok/s est.) is the previous generation of the top pick and leaves about 5GB for KV-cache. GPT-OSS 20B (ollama run gpt-oss:20b, MXFP4 ~13.8GB, ~85 tok/s est.) stays a fast chat option. For dense 27B work beyond Qwen3.8, Qwen3.5 27B Instruct (~16GB) and Qwen3.6 27B (~18GB) run at the same ~23 tok/s est.

What 36GB Cant Run

The binned chip gives up capacity as well as bandwidth. Several 35B-class models fit their weights but leave little room for context:

• Laguna XS 2.1 (33B MoE, ~20.3GB) fits weights tightly, with about 5GB left for KV-cache.
• Ornith 1.0 35B (~21.2GB at Q4) is a 35B MoE that leaves almost no room for context at 36GB.
• Gemma 4 31B (~20GB) fits weights, same story.

If your work targets 70B-class models or 120B MoE, the M5 Ultra is your real target.

FAQ

Can the Mac Studio M5 Max 36GB run a 35B model?

Yes, as long as it is an MoE. Qwen3.6 35B-A3B, the top pick, loads in ~22GB at Q4_K_M and leaves about 3GB for KV-cache; Qwen3.5 35B-A3B Instruct (~20GB) leaves about 5GB. Ornith 1.0 35B (~21.2GB) fits the same way, with little room for long contexts.

How fast is the 36GB config for LLM inference?

The binned 460 GB/s bus delivers interactive speeds. MoE models reach ~84 to ~141 tok/s est. on 24B-35B models. Dense 27B models run at ~23 tok/s est., comfortable for chat and fine for batch tasks.

Should I buy 36GB or save for 48GB?

36GB is correct if you plan to run 27B dense or 26B-35B MoE models at Q4. Buy 48GB if you want the full 614 GB/s bandwidth, need room for 35B MoE with long context, or want future-proofing for next-generation models.

Is 36GB enough for coding with local AI?

Yes. Qwen3.8 27B at ~23 tok/s est. is a strong coding assistant, and GPT-OSS 20B at ~85 tok/s est. handles quick completions well. The MoE picks with their fast decode make multi-step agent workflows snappy.