Local LLMs on Mac: key numbers
According to the LLMCheck index (17 October 2026), MiMo-V2.6-Flash ranks #1 (Score 82; ~79 tok/s estimated on an M5 Ultra) but needs a Mac Studio with 192 GB or more. On a Mac with 32 GB or less, the best pick is Qwen 3.6-35B-A3B (Score 81; 64 tok/s community-measured on a Mac mini M6 32 GB).
Quick answers
What is the best local LLM for a Mac?
According to the LLMCheck index, as of 17 October 2026, MiMo-V2.6-Flash ranks #1 (LLMCheck Score 82) at ~79 tok/s estimated on an M5 Ultra, but it needs a Mac Studio with at least 192 GB of unified memory. On a Mac with 32 GB or less, the best pick is Qwen 3.6-35B-A3B (Score 81; 64 tok/s community-measured on a Mac mini M6 32 GB; needs 24 GB).
What is the fastest local LLM on a Mac?
The fastest measured model is Maple Preview 20B-A1B at 387 tok/s on an M5 Pro (published, Ternary); its ternary weights decode through compute-bound kernels, so among bandwidth-bound models the fastest measured reference figure is LFM2.5-2.6B at 220 tok/s published on an M5 Max (Q4_K_M).
#1 overall, and the best pick at 32 GB or less
MiMo-V2.6-Flash ranks #1 (LLMCheck Score 82) at estimated~79 tok/s estimated on an M5 Ultra (MXFP4), but it needs a Mac Studio with at least 192 GB of unified memory. On a Mac with 32 GB or less, the best pick is Qwen 3.6-35B-A3B (Score 81; community~64 tok/s community-measured on a Mac mini M6 32 GB; needs 24 GB).
Strongest dense model
Qwen3.8-27B — capability 46/50, 19 GB minimum, estimated~29 tok/s estimated on an M5 Max (Q4_K_M).
Fastest measured model
Maple Preview 20B-A1B at sourced387 tok/s on an M5 Pro (published, Ternary, MLX). Among bandwidth-bound models, the fastest measured reference figure is LFM2.5-2.6B at sourced~220 tok/s published on an M5 Max (Q4_K_M).
Best pick for each RAM tier
Each tier’s pick is the best-ranked model that fits the tier, reaches 15 tok/s on the tier's representative Mac and has a standard Mac runtime (the /best-llm/by-ram/ pick).
| Tier | Best pick | Needs | tok/s | On | |
|---|---|---|---|---|---|
| 8 GB | Nanbeige4.2-3B | 2 GB | estimated~30 | MacBook Air M1 8 GB | by RAM |
| 16 GB | Gemma 4 12B | 9 GB | estimated~17 | Mac mini M6 16 GB | by RAM |
| 24 GB | Qwen 3.6-35B-A3B | 24 GB | community~64 | Mac mini M6 24 GB | by RAM |
| 32 GB | Qwen 3.6-35B-A3B | 24 GB | community~64 | Mac mini M6 32 GB | by RAM |
| 64 GB | Qwen 3.6-35B-A3B | 24 GB | estimated~92 | Mac mini M5 Pro 64 GB | by RAM |
| 128 GB | Qwen3.8-Flash-Next | 120 GB | community~44 | Mac Studio M5 Max 128 GB | by RAM |
| 192–256 GB | MiMo-V2.6-Flash | 170 GB | estimated~79 | Mac Studio M5 Ultra 256 GB | by RAM |
| 512 GB | MiMo-V2.6-Flash | 170 GB | estimated~79 | Mac Studio M5 Ultra 512 GB | by RAM |
Largest model on one Mac
Muse Spark Open (816B MoE, 42B active): sourced21 tok/s on a 512 GB M3 Ultra Mac Studio (published, oQ3e). The largest models that fit one Mac are Kimi K2.5, Kimi K2.6, Kimi K2.7 Code (1T MoE), from 417 GB at about 3-bit; no fully specified Mac run of them is published yet — a community report for Kimi K2.7 Code (about 21 tok/s community-reported on a 512 GB M3 Ultra, in the Inferencer app; model card) does not state its decode mode or the app version, so the index does not list it.
Measured vs estimated
383 speed rows: sourced37 published by vendors or press, community64 linked community runs, estimated282 estimates from the published bandwidth model. The 101 measured rows cover 37 models (every one is listed in the measured table); for 24 of the 111 models the leaderboard figure itself is measured, and the rest show an estimate on their reference chip.
Apple Silicon memory bandwidth by chip
| Chip | GB/s | Source |
|---|---|---|
| M1 | 68 | |
| M1 Pro | 200 | |
| M1 Max | 400 | |
| M1 Ultra | 800 | |
| M2 | 100 | |
| M2 Pro | 200 | |
| M2 Max | 400 | |
| M2 Ultra | 800 | |
| M3 | 100 | |
| M3 Pro | 150 | |
| M3 Max | 400 | |
| M3 Ultra | 819 | Apple: over 800 GB/s |
| M4 | 120 | |
| M4 Pro | 273 | |
| M4 Max | 546 | |
| M5 | 153 | |
| M5 Pro | 307 | Apple: up to 307 GB/s |
| M5 Max | 614 | Apple: up to 614 GB/s |
| M5 Ultra | 1,228 | Apple: 1.2 TB/s, 50% above M3 Ultra |
| M6 | 170 | Apple: up to 170 GB/s |
Binned configurations carry their own figure: M5 Max 36 GB at 460 GB/s; M6 16 GB at 153 GB/s.
M5 Ultra vs M3 Ultra, decode
Apple ML Research decode runs on one Mac Studio, M5 Ultra vs M3 Ultra: Mistral Medium 3.5 FP8: sourced7.7 vs sourced4.95 tok/s (1.56×); Qwen3.8-27B BF16: sourced19.69 vs sourced12.46 tok/s (1.58×). The bandwidth ratio alone (1,228 vs 819 GB/s) predicts 1.50×.
First press M6 figures
The index's first press-published Mac mini M6 figures: Qwen 3 30B-A3B sourced62.01 tok/s (quant unstated, MLX); Qwen 3 4B sourced53.35 tok/s (quant unstated, MLX), published 21 September 2026.
Common misconceptions
Claims about Macs and open models that are often repeated and are wrong, each with a primary source.
Is there an M4 Ultra Mac Studio?
Myth: There is an M4 Ultra.
No. Apple's March 2025 Mac Studio shipped with the M4 Max and the M3 Ultra; Apple never made an M4 Ultra. The next Ultra chip is the M5 Ultra (August 2026, 1,228 GB/s).
What is the M5 Max memory bandwidth?
Myth: The M5 Max has 600 GB/s and the M5 Pro 273 GB/s (a figure used at the time by this index too, until 5 October 2026).
Apple gives up to 614 GB/s for the M5 Max and up to 307 GB/s for the M5 Pro. The 32-core-GPU M5 Max is a binned part at 460 GB/s.
Are there open weights for Llama 5, Qwen 4.1 or Phi-5?
Myth: Llama 5, Qwen 4.1 and Phi-5 are open-weight models you can download.
No. As of 17 October 2026 none of them has weights on Hugging Face: Meta's meta-llama organisation has no Llama 5 repository, Qwen's organisation has no Qwen 4 or 4.1 repository, and Microsoft's newest Phi models are the Phi-4 family.
Can I download Qwen 4?
Myth: Qwen 4 has been released as open weights.
Not yet. Qwen 4 has been announced, but as of 17 October 2026 the newest language models in the Qwen Hugging Face organisation are the Qwen3.8 family (August 2026). The index adds a model only once its weights and card exist.
Is Muse Spark Open the same as Muse Spark 1.2 or 1.3?
Myth: Muse Spark Open is Muse Spark 1.2 (or 1.3) with its weights released.
No. Muse Spark Open (816B total, 42B active, 15 Oct 2026) shares its base model with Muse Spark 1.3 but has additional safety post-training; Meta has not released Muse Spark 1.2 or 1.3 weights.
Are MTP speeds comparable to normal decode speeds?
Myth: A speed measured with MTP or speculative decoding is the model's decode speed.
No. Multi-token prediction (MTP) and speculative decoding measure a different workload. Nemotron 3.5 Super VL decodes at 30.6 tok/s on an M5 Max with MTP off and 57.1 tok/s with its MTP head on. The index records only plain-decode figures and cites MTP results in prose.
Does LM Studio run on Intel Macs?
Myth: LM Studio has an Intel Mac build.
No. LM Studio's system requirements say Intel-based Macs are not supported; it needs Apple Silicon.
Does Ollama use MLX on a Mac?
Myth: Ollama always runs models through llama.cpp.
Yes — since Ollama v0.40.0 (25 September 2026), models whose architecture the MLX runtime supports run on MLX by default on Apple Silicon; other models still run through llama.cpp.
Was KAT-Coder-V2.5's #1 ranking based on a measured speed?
Myth: KAT-Coder-V2.5 topped the index on a measured speed.
No. Its 5 October 2026 #1 rested on an estimate of ~117 tok/s on an M5 Max. The first linked run measured 89.3 tok/s (community-measured, OptiQ-4bit); on that figure it ranks #3 as of 17 October 2026.
Quote any block with its anchor link and the date shown. How to cite LLMCheck → · Download the data → · Methodology →