Open-Weight Models Worth Running in November 2026: A Shortlist for NZ and Australian Businesses

This shortlist is dated November 2026 and we will refresh it. Open-weight releases now land every few weeks, and a list that is three months old is already missing something. Every model below was checked on its own Hugging Face page on 9 November 2026, and the licence and size are given as the developer states them there.

We do not rank them. A benchmark score tells you how a model did on someone else's test; it does not tell you how it will do on your quotes, your claims or your contracts. The section at the end explains how to choose between them on your own work, which is the only ranking that counts.

First, the two filters that matter more than the model

Licence. Open weight means you can download and run the model. What you may do with it is set by the licence. Apache 2.0 and MIT let you use, change and sell on top of the model with few conditions. A custom licence can add conditions, often aimed at very large companies. Read the licence file, not a summary of it.

Hardware. A model you cannot run is not on your shortlist. The size that matters is the total parameter count, because the whole model has to sit in memory even when only part of it works on each word. A "mixture of experts" model with 501B total and 23B active still needs the memory for 501B; it just runs faster than a dense model that size.

So the shortlist is grouped by what it takes to run.

Runs on one workstation or a high-memory Mac

These are the models a small business can run in its own office, with no data leaving the building.

Qwen3.8-27B (Alibaba's Qwen team). 27B parameters, dense. Licence: Apache 2.0. Context: 262,144 tokens natively. Hugging Face. A dense model this size is within reach of a single high-memory machine once quantised, and the licence puts no conditions on commercial use beyond Apache's usual notices.

gpt-oss-20b (OpenAI). About 21B parameters. Licence: Apache 2.0. Hugging Face. Released in August 2025, so older than everything else here, but it is still OpenAI's latest open-weight release and it is a dependable, well-supported baseline.

Nemotron 3.5 Bolt (NVIDIA). Two sizes: Bolt E4B, 8B parameters (4B effective), built on Google's Gemma 4 for fast local assistants and voice; and Bolt 74B-A9B, 74B total with 9B active, a vision-language model built for everyday agent work locally. Licence: OpenMDW-1.1, and NVIDIA's cards say both are ready for commercial use. Bolt 74B-A9B on Hugging Face. NVIDIA also publishes a 24GB build of the 74B model, which tells you the class of machine it is aimed at.

If you want to see what this tier is like in practice, our Mac mini notes cover running models on a desk-sized machine.

Needs one serious server

gpt-oss-120b (OpenAI). About 117B parameters. Licence: Apache 2.0. Hugging Face. The step up from the 20b, still on a permissive licence, and runnable on a single high-end server.

Beam 1.0 (Reflection). 501B total, 23B active. Licence: Apache 2.0. Context: 1 million tokens. Hugging Face. Released in October 2026. A large model on a fully permissive licence is still unusual, which is what puts it on this list. It needs a multi-GPU server; quantised builds for llama.cpp, Ollama and LM Studio are linked from its card.

Data-centre scale

These are for businesses with a data-centre partner or a serious hardware budget, or for a provider hosting a model for many clients.

GLM-5.3 (Zhipu, published as zai-org). Around 750B parameters by Hugging Face's count of the weights. Licence: the GLM-5.3 License, a permissive licence with one added condition: a company running a model-as-a-service business with revenue over 10 billion US dollars must pass Z.AI's security review before commercial use. Hugging Face. Its smaller sibling, GLM-5.3-Flash, is 320B total with 18B active, under MIT. Hugging Face.

Kimi K3.1 Pro (Moonshot AI). 2.8T total, 104B active, 1 million token context, published 2 November 2026. Licence: the Kimi K3.1-Pro License, permissive except that a model-as-a-service business with revenue over 20 million US dollars needs a separate agreement with Moonshot before commercial use. Hugging Face. Few businesses in New Zealand or Australia will run this themselves; it matters because local hosting providers may.

What is not here, and why

Fine-tunes and re-quantisations of the models above are not listed separately; start from the original and its licence. Models whose licence we could not read on the developer's own page are left out. And we have not included claims about which model is "best" at coding, writing or reasoning, because we cannot cite a test that would hold for your work.

How to choose between them

1. Filter by licence. If you will build a product on the model, start with Apache 2.0 or MIT and read anything custom line by line.
2. Filter by hardware. Decide where it will run: a machine in the office, a server you own, or a New Zealand or Australian data centre. That rules most of the list in or out.
3. Take twenty real examples of one job, with the answers you would accept. A month of support emails, a set of quotes, the contract clauses you check every week.
4. Run them through two or three candidates, and through the hosted model you use now, such as Claude or ChatGPT. Score the outputs side by side, on quality and on what each would cost to run at your volume.

That afternoon tells you more than any leaderboard. Often the answer is that a hosted model stays the right tool for most work and an open-weight model takes the narrow, private, high-volume job. Do you need your own AI model? covers that decision.

Running one for real

Testing on a laptop is easy. Running a model that a team depends on means hardware sized for the load, monitoring, updates when a better model ships, and someone accountable when it is slow. Private LLM infrastructure is how we approach that, and Open-weight models is where we publish what we learn. We will update this shortlist as the field moves; check the date at the top before you rely on it.