The Best Local LLM in 2026 (Picks for Every GPU and Mac)

The best local LLM is the strongest open model that actually fits in your memory with room left over for context. That is the whole trick. A great model that spills out of VRAM crawls; a smaller one that fits answers instantly. So instead of one winner, here are the picks for each hardware tier, with the exact Ollama tag to pull.

New to running models yourself? Start with how to run an LLM locally, then come back here to pick the model.

Quick picks by memory (October 2026)

Sizes are the default Ollama tags as listed in the Ollama library on October 19, 2026. They are the weights only; budget a few extra gigabytes for the context window.

8 GB: qwen3.5:4b

On an 8 GB card (RTX 4060, RTX 3070) or a base Apple Silicon Mac, qwen3.5:4b is the smallest Qwen3.5 that feels like an assistant rather than autocomplete. It is Apache 2.0 and has a 256K native context window. The 9B sibling technically squeezes in, but leaves almost no room for context.

12 GB: qwen3.5:9b

qwen3.5:9b is the sweet spot for 12 GB cards like the RTX 3060 12 GB or RTX 4070, and for Macs with 16 GB of unified memory. Its model card lists MMLU-Pro 82.5 and GPQA Diamond 81.7, which is remarkable for a model this small. If you also want image input, gemma4:e4b is the multimodal alternative at a similar size.

16 GB: gpt-oss:20b

OpenAI’s open-weight gpt-oss:20b is a mixture-of-experts model: 21B parameters in total but only about 3.6B active per token, stored in a compact format so the Ollama download is 14 GB. That is why it fits a 16 GB card and still runs fast. It is the best all-rounder at this tier, with a 128K context window and an Apache 2.0 license. gemma4:12b (about 8 GB) is the pick if you want vision.

24 GB: qwen3.6:27b

This is where a local model can replace a chat subscription for most everyday work. qwen3.6:27b is a 27B dense model with a 256K native context and an 18 to 19 GB download, so it fits an RTX 3090 or 4090, or a Mac with 32 GB of unified memory. Gemma 4 26B and 31B are the closest competitors and add image input.

32 GB and up: qwen3.6:35b or glm-4.7-flash

With an RTX 5090 or a 48 GB+ Mac, pull qwen3.6:35b, a 35B mixture-of-experts model with only about 3B parameters active per token, which keeps it fast. glm-4.7-flash is a near twin in size (about 19 GB at Q4) and ships under the MIT license if you prefer that. Beyond a single card, gpt-oss:120b is a 65 GB download and needs serious memory.

Best local LLM for coding

qwen3-coder:30b is the coding model to beat locally. It is a 30B mixture-of-experts model with around 3.3B active parameters, a 256K context window (enough to hold a real chunk of a codebase) and a 19 GB download. Point your editor’s AI extension or a terminal coding agent at Ollama’s local API and you have a private coding assistant with no per-token bill. On 16 GB, use gpt-oss:20b instead.

Best for reasoning and vision

• Reasoning: deepseek-r1:14b or deepseek-r1:32b (MIT) think step by step before answering. On tiny hardware, phi4-mini-reasoning is the smallest one worth running.
• Vision: the Gemma 4 family (gemma4:e4b, gemma4:12b, gemma4:26b) reads images as well as text, under Apache 2.0.

How to pick the right one

• Fit first. Download size plus a few GB for context must fit in VRAM, or in unified memory on a Mac (leave room for macOS too).
• Mixture-of-experts is your friend. MoE models like gpt-oss, qwen3.6:35b and qwen3-coder only run a slice of their weights per token, so they are much faster than their size suggests.
• On a Mac, try the MLX tags. Ollama lists -mlx variants (for example qwen3.6:27b-mlx) built for Apple Silicon.
• Check the license if you ship a product. Everything above is Apache 2.0 or MIT.

Try one in a minute

Install Ollama, then run ollama run qwen3.5:9b (swap in the tag for your tier). That pulls the model and drops you into a chat. From there you can connect a web UI or wire the local API into your own self-hosted AI stack, ideally on a dedicated home server.

Model tags change every few weeks, and benchmark charts rarely match what runs well on your exact machine. Inside the Tinkerer Club, members post which models and settings actually work on their GPUs, Macs and homelabs, so you can skip the weekend of trial and error.