The best local LLM for a Mac mini M4 Pro (48 GB) is KAT-Coder-V2.5 at ~66 tok/s (estimated), the top-ranked model that reaches 15 tok/s here. Its 48 GB fits 71 of the 130 models in the LLMCheck index, up to 80B-class. The most capable, Qwen3.8-27B, runs at ~13 tok/s (estimated): slow for chat.
Considering this Mac?
Mac mini M5 Pro (48GB): build-to-order from Apple · Mac mini M4 Pro on Amazon Renewed → · compare every Mac for local AI →
Ranked by LLMCheck suitability (capability balanced against speed on the M4 Pro); the top pick is the best-ranked model that reaches 15 tok/s here, has a standard Mac runtime and fits the GPU memory limit (raised if need be), and slower models are tagged slow. Click a model for its full figures and setup. Every speed carries its provenance mark: most are index estimates from the memory-bandwidth model, a filled square is a published figure and a hollow circle a community run — submit a real run →
| # | Model | Size | License | Speed | 64K context | Capability |
|---|---|---|---|---|---|---|
| 1 | KAT-Coder-V2.5 | 35B MoE | Apache 2.0 | estimated66 tok/s | agent-ready | 39/50 |
| 2 | Qwen3.8-27B slow | 27.8B | Apache 2.0 | estimated13 tok/s | agent-ready | 46/50 |
| 3 | Nanbeige4.3-33B slow | 33.1B | Apache 2.0 | estimated11 tok/s | agent-ready | 46/50 |
| 4 | Qwen 3.6-35B-A3B | 35B MoE | Apache 2.0 | estimated74 tok/s | agent-ready | 38/50 |
| 5 | Recursive Flint | 39.8B MoE | Apache 2.0 | community80 tok/s | agent-ready | 37/50 |
| 6 | LFM2.5-24B-A2B | 24B MoE | LFM Open | estimated90 tok/s | agent-ready | 35/50 |
| 7 | Qwen 3.6-27B slow | 27B | Apache 2.0 | estimated14 tok/s | agent-ready | 44/50 |
| 8 | Nemotron 3.5 Lightning | 30B MoE | OpenMDW | estimated67 tok/s | agent-ready | 36/50 |
| 9 | MiniCPM-V 4.7 35B-A3B | 35.2B MoE | Apache 2.0 | community75 tok/s | agent-ready | 35/50 |
| 10 | North Mini Code 1.0 | 30.5B MoE | Apache 2.0 | estimated65 tok/s | agent-ready | 36/50 |
| 11 | Gemma 4 26B-A4B | 26B MoE | Apache 2.0 | estimated72 tok/s | agent-ready | 35/50 |
| 12 | Nemotron 3.5 Bolt 74B-A9B raise GPU limit | 74B MoE | OpenMDW | estimated28 tok/s | 40/50 |
Showing the top 12 of 71 models that fit in 48 GB. See the full leaderboard or all benchmarks.
Context: “fits” means the weights at a short prompt. Coding agents (Claude Code, OpenCode, Codex CLI) hold 64K tokens or more, and every token adds KV cache: agent-ready marks a model whose weights plus a 64K-token 16-bit KV cache still fit in the 42 GB this Mac’s GPU can use with the limit raised (57 of the 66 models here with a KV figure). That is an estimate from each model’s config.json, not a measurement; a q8 KV cache halves the cache part, and — means the index has no KV figure for the model yet. The GPU memory limit below applies on top.
GPU memory limit: macOS lets the GPU use about 36 GB of this Mac’s 48 GB by default on macOS 15 and earlier, and about 37.4 GB on macOS 26 and later. Of the 71 models that fit, 7 need more than that at a short prompt: 3 need a raised limit and are tagged raise GPU limit (Nemotron 3.5 Bolt 74B-A9B and Llama 3.3 70B only past the safe setting below, as measured runs on this much memory show); 4 need more than even the raised limit and are tagged over GPU limit. sudo sysctl iogpu.wired_limit_mb=43008 raises it to 42 GB, leaving 6 GB for macOS, until the Mac restarts (sudo sysctl iogpu.wired_limit_mb=0 restores the default). How the limit is read →
The index has no verified Ollama tag for KAT-Coder-V2.5, so start from its listing and copy the exact name there: Search the Ollama library → · model card →
GGUF builds run in LM Studio or llama.cpp; MLX builds run in LM Studio or mlx-lm.
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 M4 Pro page for its figures and setup.
Twice as fast: no Mac in the index runs KAT-Coder-V2.5 at twice this Mac’s estimated~66 tok/s; the fastest figure is the Mac Studio M5 Ultra (96 GB) at estimated~120 tok/s, $5,499 new (Apple list price, checked 2026-11-14): build-to-order from Apple (Amazon does not stock this configuration).
For less: the cheapest Mac that still runs KAT-Coder-V2.5 at 15+ tok/s is the MacBook Air M2 (24 GB) at estimated~31 tok/s, about $850 refurbished or used: search Amazon Renewed listings →. New: the Mac mini M6 (24 GB) at estimated~47 tok/s, $1,099 new (Apple list price, checked 2026-11-14): Mac mini M6 (24GB) on Amazon →.
Speeds are each Mac’s own figure from the index, with its provenance mark. Prices come from data/macs.json: Apple’s list price for a new Mac (dated), an approximate street price for a used or refurbished one. Amazon stock checked 2026-11-18. Amazon links are affiliate links: as an Amazon Associate, LLMCheck earns from qualifying purchases.
Apple no longer sells the Mac mini M4 Pro (48 GB) new. A current Mac mini M5 Pro (48 GB) runs the same models with 307 GB/s of memory bandwidth instead of 273 GB/s, so they generate faster. The Mac mini M5 Pro (48 GB) is build-to-order from Apple; Amazon does not stock that configuration.
As an Amazon Associate, LLMCheck earns from qualifying purchases. Affiliate links cost you nothing extra and never influence our rankings.
LLMCheck. (2026). Best Local LLMs for the Mac mini M4 Pro (48 GB). llmcheck.net. https://llmcheck.net/best-llm/mac-mini-m4-pro-48gb/ (updated ). CC BY 4.0.
KAT-Coder-V2.5 (35B MoE, Apache 2.0) is the best all-round pick at ~66 tok/s (estimated) on the Mac mini M4 Pro. If you want maximum speed, LFM3-1.7B hits ~198 tok/s (estimated); Qwen3.8-27B is more capable but runs at only ~13 tok/s (estimated) here.
About 71 of the 130 models in the LLMCheck index fit in 48 GB of unified memory, from compact models up to Qwen3-Coder-Next (80B MoE).
Just. A dense 70B such as Llama 3.3 70B needs about 44 GB at 4-bit, which leaves little of the 48 GB for context and other apps; the index puts it at ~5 tok/s (estimated) here.
48 GB is enough for dense models up to about 72B at 4-bit (the largest in the index that fits is Qwen 2.5 72B), and for mixture-of-experts models such as Nemotron 3.5 Bolt 74B-A9B (~28 tok/s (estimated)). As a rule of thumb a 4-bit model needs about 0.57 GB per billion parameters plus ~2 GB, and macOS needs its own share of memory too. Because Apple Silicon uses unified memory, the same pool serves as system RAM and GPU memory.