Best Local LLM by Task in 2026, and the Mac mini That Runs It

Cloud Mac · By FideNubi Team · 2026-10-30 · Updated 2026-11-15

Short answer. There is no single best local AI model: pick a model built for your task, then a Mac mini with enough memory for it. For coding, the specialist models (Qwen3-Coder, Devstral Small 2, North Mini Code, Ornith 1.5 35B, Flint) need a 32 GB Mac mini, while Ornith 1.5 9B runs on 16 GB. For AI agents such as OpenClaw, gpt-oss 20B needs 24 GB and NVIDIA's Nemotron 3.5 Lightning 48 GB. For documents and OCR, GLM-OCR and the small Qwen3-VL sizes run on 16 GB, and so do translation (TranslateGemma 4B and 12B) and a search-your-documents setup (Qwen3 Embedding plus a small chat model). Images (FLUX, Z-Image, Krea 2) run on 16–24 GB, speech to text and text to speech run on any Mac mini, music (ACE-Step) wants 24 GB, 3D works only through community ports, and video runs but slowly. Choose a task below to see every model for it and the cloud Macs that can run them.

Key facts

• Our catalog has 46 local models: 27 language models (58 sizes) plus 19 for images, video, speech, music and 3D, checked 2 November 2026. We review the list every month. We do not rank models by quality and have not benchmarked them; speeds for non-text models are measurements reported by developers and users, with links.
• Image and video generation on Apple silicon is limited by compute, not memory bandwidth: the M6 and M5 Pro, whose GPUs have neural accelerators, are much faster than an M4 Pro (Draw Things and vpipe measurements).
• Coding specialists: Qwen3-Coder 30B, Devstral Small 2, North Mini Code 1.0, Ornith 1.5, Flint. Cheapest cloud Mac with 32 GB at our latest check: from €129 a month (MacDuty, M4 32 GB).
• Agent models: Nemotron 3.5 Lightning (NVIDIA names OpenClaw and Hermes Agent), Muse Glimmer (Meta), gpt-oss (OpenAI), LFM3 1.7B (Liquid AI, post-trained for OpenClaw and Hermes Agent), plus the coding agents above.
• Documents and OCR: GLM-OCR (documents only), Qwen3-VL (images, charts, screens); both have sizes that run on 16 GB.
• A Mac mini goes up to 32 GB with the M4 or M6 and 64 GB with the M4 Pro or M5 Pro, so no Mac mini runs gpt-oss 120B or Qwen 3.5 122B.
• Memory figures are for 4-bit (Q4) builds with a 32K-token context; the formula is explained in Which Mac mini for a local LLM?.

Best local AI models by task, and the Mac mini that runs them

Choose a task. The table lists every model whose developer names that use, the Mac mini its smallest size needs, the largest size a Mac mini can run, and the cheapest cloud Mac for it. "Details" opens the model: memory breakdown, the Mac mini configurations that fit with estimated speed, all offers that can run it (cheapest first) and the command to install it.

What is the best local LLM for coding?

Two kinds of models are worth trying. Coding specialists are trained for software-engineering agents: Qwen3-Coder (30B with 3.3B active parameters and a 256K context for whole repositories), Devstral Small 2 (Mistral; explores codebases with tools and edits several files), North Mini Code 1.0 (Cohere; trained across agent harnesses such as SWE-Agent and OpenCode) Ornith 1.5 (trained with a self-improvement loop that generates its own coding tasks) and Flint (Recursive; investigates bugs, implements features and writes tests across many tool calls). General models with strong coding, such as Qwen 3.8, Qwen 3.6 (the developer highlights agentic coding), Gemma 4, gpt-oss and the DeepSeek-R1 distills, are useful when the same assistant also writes and reasons.

Mixture-of-experts models (Qwen3-Coder, North Mini Code, Ornith 1.5 35B, Flint) read only about 2.5 to 3 billion parameters per token, so on the same Mac they answer several times faster than a dense 24B model such as Devstral Small 2.

What is the best local LLM for AI agents like OpenClaw?

An agent model has to call tools reliably, keep track of a multi-step task and recover from errors. NVIDIA built Nemotron 3.5 Lightning "for the execution layer of always-on agents" and names OpenClaw and Hermes Agent; Meta built Muse Glimmer for autonomous agents on consumer hardware, fully offline; OpenAI designed gpt-oss for agentic tasks, function calling and Python tool calls. The coding agents above work for developer agents. On a 16 GB Mac mini, the small Qwen 3.5, Gemma 4 and Granite 4.2 sizes support tool calling, as does NVIDIA's compact Nemotron 3.5 Bolt; Liquid AI post-trained the tiny LFM3 1.7B (about 1 GB at 4-bit) for agent harnesses such as OpenClaw and Hermes Agent.

An always-on agent runs around the clock, which is where a rented Mac mini helps; see Mac mini for OpenClaw and AI agents: rent, buy or VPS?.

Which local LLM is best for documents and OCR?

GLM-OCR (Z.ai) does one thing: it reads complex documents (text, tables, formulas, figures) and extracts information; its developer reports first place on the OmniDocBench V1.5 benchmark. Qwen3-VL understands images, screenshots, charts and video and can answer questions about them. For summarizing long text, Meta lists summarization for Llama 3.1 8B and retrieval and summarization for Llama 3.2.

Which local LLM can understand images?

Many current models take images as input: Gemma 4, Qwen 3.5 and Qwen 3.8, Mistral Small 3.2, Muse Glimmer, Devstral Small 2 (for screenshots of code and interfaces), Qwen3-VL, GLM-OCR and Gemma 3n. Pick by size: the small Gemma 4 and Qwen 3.5 sizes and Gemma 3n run on 16 GB, the 24–30B models need 32 GB. Choose "Images" in the tool above to see them all.

Which local LLM is best for reasoning and math?

Reasoning models think step by step before answering, which costs time and context. The DeepSeek-R1 distills (1.5B to 70B) are built for mathematics, programming and logic; gpt-oss and Qwen 3.8 have adjustable reasoning; Phi-4 (Microsoft, 14B) focuses on reasoning in English for memory-constrained setups but has a short 16K context. Choose "Reasoning and math" in the tool above.

Which local LLM is best for chat and writing?

For an everyday assistant, the general families cover every Mac mini: Gemma 4 and Qwen 3.5 from tiny to 31–35B, Llama 3.1 8B and Llama 3.3 70B (Meta; the 70B fits only the 64 GB Mac mini, tightly), Mistral Small 3.2 and Granite 4.2. For many languages, Gemma 3n was trained on data in over 140 languages and Qwen 3.5 highlights broad language coverage.

Which local LLM is best for translation?

TranslateGemma (Google, built on Gemma 3) is made only for translation between 55 languages and can also translate text in images. The 4B and 12B sizes run on a 16 GB Mac mini; the 27B needs 32 GB.

How do you search your own documents with a local LLM (RAG)?

Retrieval-augmented generation (RAG) uses two models. An embedding model such as Qwen3 Embedding turns your documents into vectors so the right passages can be found; it does not chat. A chat model then answers from those passages; IBM built Granite 4.2 with RAG, tool use and structured JSON output in mind. Run both on the same Mac and add their memory needs.

Which local model can generate images on a Mac mini?

Image models run well on Apple silicon through the free Draw Things app or the command-line tool mflux (MLX). Z-Image-Turbo, FLUX.2 [klein] 4B, Krea 2 Turbo and SDXL work on a 16 GB Mac mini; FLUX.1 [dev] and the 4-bit Qwen-Image 2.1 want 24 GB. Reported speeds: about 50 seconds for a 1024×1024 FLUX.1 [dev] image on a 24 GB Mac mini M4 (blog test), about 35 seconds for Krea 2 Turbo on an M5 Pro with 24 GB (vpipe developer), but around 5 minutes for Qwen-Image on an M4 Pro. Check the licence: FLUX.1 [dev] and Qwen-Image are not licensed for commercial use.

Can a Mac mini generate video?

It can, slowly. The open video models are built for NVIDIA GPUs; on a Mac they run through Draw Things or vpipe. The vpipe developer measured a 5-second MiniMax H3 clip (960×544) at 17 minutes 43 seconds on a Mac mini M4 Pro with 64 GB and 5 minutes 29 seconds on an M5 Pro; a Reddit user generated LTX-2.5 at about one minute per second of low-resolution video on an 8 GB MacBook Air; Wan 2.2 took 82 minutes for 2 seconds on an M1 Max. For regular video work, a cloud GPU is usually the faster choice; what a clip costs on a cloud Mac compared with video services is in Cloud Mac mini for AI video: cost per clip vs services.

Which local model is best for speech to text?

Transcription is fast on every Mac mini. Whisper large-v3-turbo (OpenAI, 99+ languages) runs through whisper.cpp or mlx-audio; Parakeet TDT 0.6B v3 (NVIDIA, 25 European languages) through parakeet-mlx is faster; Qwen3-ASR covers more Asian languages. A Reddit user who benchmarked the engines on an M2 Pro found all of them far faster than real time, and macOS 26's built-in speech engine more accurate than Whisper Small for English.

Which local model is best for text to speech?

Kokoro-82M is tiny and fast, Qwen3-TTS adds voice cloning and voice design, Chatterbox speaks 23 languages with cloning, and Phonon runs directly in Ollama (ollama run phonon "Hello"). All of them fit any Mac mini; most run through mlx-audio.

Can a Mac mini generate music?

ACE-Step 1.5 (MIT licence) writes full songs with vocals and ships official macOS scripts that use MLX. Its README recommends 20 GB or more, and users report crashes on 16 GB, so plan for a 24 GB Mac mini. One user generated two 100-second songs in 58 seconds on an M1.

Can a Mac mini create 3D models?

Only through community ports: the original TRELLIS.2 (Microsoft) and Hunyuan3D 2.1 (Tencent) depend on NVIDIA-only code. The trellis-mac port turns one photo into a textured mesh in about 5 minutes on an M4 Pro with 24 GB; the Hunyuan3D port needs about 6 minutes for the shape on the same Mac, and 24 GB is tight.

How do you install a model on a cloud Mac?

Connect to the cloud Mac, then follow the steps the tool shows for the model. Language models and Phonon use Ollama (curl -fsSL https://ollama.com/install.sh | sh or the installer from ollama.com; macOS Sonoma 14 or newer), for example ollama run qwen3-coder:30b; image and video models use Draw Things, mflux or vpipe; speech models use whisper.cpp, parakeet-mlx or mlx-audio. Already have a Mac in mind? What LLM can a Mac mini run? lists every model per memory size. The step-by-step guide, how we calculate memory and speed, and Apple's memory bandwidth for each chip are in Which Mac mini for a local LLM?. Every plan we track is in the price tables for Europe, North America and Asia and Australia.

← All articles