The open-weight models worth knowing in late 2026
Size, licence and hardware notes for the open-weight model families worth knowing in late 2026, from DeepSeek and Kimi K3 to Gemma 4, gpt-oss and Llama 4.
There are now more open-weight models than any small business could test, and new ones arrive every few weeks. You do not need to know them all. You need to know which families exist, roughly what size they are, and what their licences let you do, so you can rule most of them out quickly.
Here are the families we think are worth knowing as of late October 2026. It is a snapshot, not a ranking. The sizes and claims below are the makers’ own, taken from their model cards, and where sources disagree we say so. If the term open weights is new to you, start with why open weight is not the same as open source.
The very large mixture-of-experts models
Most of the largest families use a mixture-of-experts design: the model holds many expert networks, but only a few run for each token. So every entry below has two numbers. The total decides how much memory the model needs, and the active count decides how much work each token takes.
- DeepSeek V4 and V4.1. DeepSeek V4-Pro has 49 billion active parameters and a 1 million token context. Its total is 1.6 trillion on the April model card and 1.7 trillion on the card for the August general release. V4-Flash is smaller, at 284 billion total with 13 billion active. V4.1-Flash, the smallest model in the newer family, also understands images. The model cards for V4-Pro-0813 and V4.1-Flash say the weights are MIT licensed. Our DeepSeek V4 post covers the Pro model in more detail.
- Xiaomi MiMo-V2.6. MiMo-V2.6-Pro has 1.02 trillion total parameters, with 42 billion active and a 1 million token context. Its card lists text, image, video and audio input. MiMo-V2.6-Flash has 309 billion total with 15 billion active. Both are MIT licensed on their Hugging Face pages. See our MiMo-V2.6-Pro post for more.
- Kimi K3. Moonshot’s K3 has 2.8 trillion total parameters, with 104 billion active and a 1 million token context. It uses its own Kimi K3 License. Commercial products with over 100 million monthly active users or over US$20 million monthly revenue must prominently display “Kimi K3”, and businesses that offer the model as a service to third parties with revenue above US$20 million over any 12 months need a separate agreement with Moonshot AI. Neither condition applies to internal use that does not expose the model to third parties.
- Qwen3.8 and GLM-5.x. Alibaba’s open Qwen3.8 checkpoint is a text-only model with 2.4 trillion total parameters and 95 billion active. It always runs in thinking mode and uses its own Qwen3.8-Max License. Zhipu’s GLM-5.2 is MIT licensed with a 1 million token context. GLM-5.3 has 744 billion total parameters with 40 billion active, and Z.ai’s card calls it “the most capable open-weights model for coding”, which is the company’s own claim. Its model card and its GitHub repository do not give a matching licence, so read the licence file in the model repository before you rely on it.
Smaller and mid-sized models
The models in this section are the ones most small businesses will test first, because their makers size them for a workstation or a well-equipped laptop.
- Gemma 4. Google’s family runs from E2B, at about 2.3 billion effective parameters, to a 31 billion parameter dense model. The E2B and E4B models accept text, image and audio, and the 26B and 31B models accept text and image. Google’s memory table, which covers the weights with a 20% overhead added, puts the 31B model at about 17.5 GB at 4-bit quantisation, and the 26B mixture-of-experts model at about 14.4 GB. The licence is Apache 2.0.
- gpt-oss. OpenAI’s gpt-oss family has two open-weight models. The 120B model has 117 billion total parameters with 5.1 billion active, and its card says it fits on a single 80 GB GPU. The 20B model has 21 billion parameters with 3.6 billion active, and its card says it can run within 16 GB of memory with MXFP4 quantisation. Both are listed as Apache 2.0.
- Qwen3.8-27B. The 27 billion parameter model in the Qwen3.8 family is Apache 2.0 licensed and accepts image and video input. It sits alongside the Qwen3.6-27B model released in April 2026, which is also Apache 2.0.
- Mistral Small 4. This model has 119 billion total parameters with 6 billion active (its card says 6.5 billion), a 256k context and text and image input. It was released on 16 March 2026 under Apache 2.0. Mistral’s model list also shows a Large 4, which we leave out until its weights and licence text are published.
- Llama 4 Scout. Meta’s Scout model has 109 billion total parameters with 17 billion active, spread across 16 experts, and a 10 million token context. It is released under the Llama 4 Community License, which is covered in the next section.
Licences range from MIT to custom
The licence decides what you may do commercially, so it is the first filter. Here is how the families above are licensed as of late October 2026, based on their model cards and licence files.
| Licence | Families in this post |
|---|---|
| MIT | DeepSeek V4-Pro and V4.1-Flash, Xiaomi MiMo-V2.6, GLM-5.2 |
| Apache 2.0 | Gemma 4, gpt-oss, Mistral Small 4, Qwen3.8-27B and Qwen3.6-27B |
| Custom | Kimi K3, the Qwen3.8 open checkpoint, Llama 4 |
| Unclear; check the licence file | GLM-5.3 |
Custom licences add conditions on top of the usual permissions: a display of the model name, a separate arrangement above a user or revenue threshold, or rules for hosted services. Most small businesses sit far below those thresholds, but read the licence file itself rather than a summary before you build on the model.
What the list shows
Three patterns stand out. First, the largest models are trillion-parameter mixture-of-experts designs that most businesses will reach through a hosting provider, not their own hardware. Second, licences range from MIT and Apache 2.0 to custom terms, so the word “open” does not tell you what you may do. Third, the smaller models are now a practical place to start. Google says Gemma 4 targets laptops, workstations and consumer GPUs, and OpenAI says gpt-oss-20b runs within 16 GB of memory.
What this means for a small business
Start from the task, not the leaderboard. Write down what the model must do, the data it will see and where that data may live. Then check the licence against how you will use the model, and only then look at hardware or a hosted price. A small model that fits the task on your own machine may serve you better than a very large one that needs a hosting provider and a longer approval process. This is the order we follow when we build AI chatbots and agents for a business.
Then shortlist two or three models from this list and run them on your own work. Ten real emails, invoices or support questions will tell you more than any published score.
Sources
Start with the people doing the work.
We build software, AI chatbots, and AI agents around the systems you already run, with people keeping the judgment.