← All articles
2026-07-027 min read

Best Open-Weight LLMs to Run at Home in 2026

The open-weight models worth downloading in September 2026: parameters, context windows and licences from the model cards, plus the four no desk can hold.

The open-weight field moves fast enough that most "best models" lists are wrong within a quarter. This one is dated on purpose, and every number in it comes from a model card rather than from memory. If you are choosing something to download in September 2026, these are the candidates, sorted by what your machine can hold.

First, what "open-weight" means

An open-weight model is one whose weights, the trained model file itself, are published for anyone to download and run. That is what makes local AI possible: the file lives on your disk, runs on your hardware, and works offline. It is not the same as open source in the strict sense, since training data and licences vary, and the licence part matters more than it used to. All five models in the table below are Apache 2.0. The ones with stricter terms come later in this piece, and they are also the ones you cannot run anyway.

You will usually download models in GGUF format, the standard container for quantized models. The everyday choice is 4-bit, which shrinks a model to roughly a quarter of its full size at a modest quality cost. If your machine's memory is the open question, read how much RAM you need for local AI first, because it decides which half of this list applies to you.

The five worth downloading

ModelSizeContextLicenceReads images
Qwen 3.8 27B27.78B dense262,144Apache 2.0yes
Qwen3-Coder 30B30.5B total, 3.3B active262,144Apache 2.0no
Gemma 4 26B25.2B total, 3.8B active256KApache 2.0yes
Mistral Small 3.2 24B24B densenot publishedApache 2.0yes
gpt-oss 20B21B total, 3.6B activenot publishedApache 2.0no

Qwen 3.8 27B, the default answer

If you download one model, this is it. Dense rather than a mixture of experts, so it fits on a single graphics card without the sparse-model caveats. It reads images and video without a separate VL variant to hunt down. Its context window is a quarter of a million tokens out of the box, and unusually, that window is affordable to actually fill. The model runs a lighter kind of attention on three layers out of four, so the conversation cache costs a quarter of what a conventional model this size would charge.

The default 4-bit build is 16.5 GB, which loads whole on a 24 GB card with room for a long conversation. Our guide to running it has the full size table for every quantization, the Mac route through MLX, and the one template setting that makes people think the weights are broken when they are not.

Qwen3-Coder 30B, for agent work

Same family, built for code and tool calling. 30.5B parameters with 3.3B active per token, which means it reads like a 30B model and generates closer to the speed of a small one. The 262,144 token context is the useful part for agents specifically, because an agent burns context on file contents and command output faster than a chat does. This is the model to point a local coding agent at.

Gemma 4 26B, the one that runs light

Google's current open model is a mixture of experts: 25.2B parameters in total with 3.8B active per token, a 256K context, and image input. The active count is the interesting number, because generation speed tracks active parameters rather than total ones. On a machine that can hold it, this reads noticeably faster than a dense model of the same size.

Mistral Small 3.2 24B, when you want it brief

The French lab's models have a distinct character: concise, quick, and free of filler. Small 3.2 is 24B parameters, Apache 2.0, and it takes image input. If you find some models exhaustingly chatty, this is the corrective. Mistral does not publish a context length on the card, so we are not quoting one.

gpt-oss 20B, for a 16 GB machine

21B parameters with 3.6B active. OpenAI's card states it runs within 16 GB of memory thanks to MXFP4 quantization, which is worth more than it looks: it is the vendor naming a memory budget, where every other figure on this page is arithmetic somebody did afterwards. If 16 GB is your ceiling, start here.

What about Llama

Meta's Llama 4 Scout is 109B parameters in total with 17B activated across 16 experts, a 10M token context, and image input. It is a genuinely interesting model and it is not a home model: 109B parameters at 4-bit is far past a consumer card. It also carries the Llama 4 Community License instead of Apache 2.0, which matters if you ship something commercial.

The older Llama 3 generation still turns up in "best local model" lists written a year ago. It has been overtaken by everything in the table above, and there is no reason to start there now.

The ones no desk holds

The current flagships are worth knowing about mostly so you stop trying to fit them.

GLM 5.3 is 753,864,139,008 parameters. Its Flash sibling sounds like the small one in the family, and its smallest usable build is still 93.1 GB, which no Mac configuration and no consumer card reaches. Kimi K3 is a 2.8T mixture of experts with a 1M context. DeepSeek's V4 line and Qwen 3.8's trillion-parameter flagships tell the same story. No quantization brings any of them to a desk.

They run in a browser tab instead. Every plan and every credit pack in LU Labs Cloud reaches the whole chat catalogue, so there is no tier to climb for the big ones, and what a Mac can run locally in 2026 works through where that line falls if Apple Silicon is what you own.

A picking guide by memory

Below 16 GB, nothing on this list fits properly. You can run something smaller, and it will show its size on anything nuanced or multi-step. Try it once to see how local feels, then decide whether the browser route suits you better.

At 16 GB, gpt-oss 20B is the one with a vendor claim behind it.

At 24 GB, Qwen 3.8 27B loads whole at 4-bit with room for context, and Gemma 4 26B and Mistral Small 3.2 both fit comfortably.

At 32 GB and up, run a heavier quantization of the same models, or keep two loaded at once: a coder for the agent and a generalist for everything else.

Two honest caveats. Best shifts every few months in this field, so treat the list as a September 2026 snapshot with a shelf life. And none of these match the largest cloud models on the hardest problems. What they match is the everyday work, privately, offline, and for free, which is the whole trade explained in local AI vs cloud AI.

The other 20% has an open-weight answer too, it just does not fit on a desk. Trillion-parameter models like Kimi K3 are open weights that nobody downloads, and they run on managed GPUs instead: Kimi K3 online is the shape that takes.

The easy way to try them

You can hunt down GGUF files and quantization levels yourself, or let an app do the matching. LU Labs is a free desktop AI studio that reads your hardware and lists only the models that fit it, then handles the download and the backend, whether that is its own LU Engine, an Ollama or an LM Studio you already run. Chat is the start; the same app carries a sandboxed coding agent and local image generation. It runs on Windows and Linux, and on a Mac as a beta for Apple Silicon on macOS 14 or newer, where local chat and Create are what has been run so far; LM Studio or Ollama stay good choices there.

The bottom line

Qwen 3.8 27B for the default, Qwen3-Coder 30B for agents, Gemma 4 26B when speed matters more than depth, Mistral Small 3.2 when you want brevity, gpt-oss 20B when 16 GB is the ceiling. All five are Apache 2.0, which is a better state of affairs than this field enjoyed a year ago. Pick the one that matches your memory and your main use, run it at 4-bit, and switch when something better lands. Switching is a download, not a commitment. The getting-started guide covers the rest.

Sources