# Which open-weight model can you actually run, and what are you legally allowed to do with it?

Date: 2026-11-02 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/which-open-weight-model-compared/

**Answer:** On one computer, the realistic options today are small models such as Gemma 4, Qwen3.8-27B, gpt-oss-20b or Nemotron 3.5 Lightning. Models above about 300 billion parameters need a server or a hosted service. Legally, Apache 2.0, MIT and OpenMDW impose few conditions, while custom licences on many flagships add rules for attribution and for businesses selling model access.

"Open" tells you one thing: you can download the file. It doesn't tell you whether your machine can hold it or whether your business may use it. Those are two separate checks, and they often point in opposite directions, because the biggest models need the most hardware and tend to carry the most conditions.

This page compares models as of November 2, 2026, using their model cards, licence files and Hugging Face file sizes. For the basics of open versus closed, see our [licence explainer](/en/learn/open-weights-vs-closed-licences/).

## Three numbers to read on any card

**Total parameters** are all the learned numbers in the file, and they largely determine the memory you need. **Active parameters** are those used per token (a small piece of text) in a **mixture-of-experts** model, where only some sections do the work at each step. They largely determine the compute, and so the speed and running cost.

The third is **bits per parameter**, set by **quantization**, which stores numbers more compactly. A trillion parameters takes about 2 TB at 16 bits, 1 TB at 8 bits, and roughly 500 GB at 4 bits. That is why a card's parameter count never tells you memory on its own. Working memory for your context comes on top of the file size too, and a million-token session can add a lot.

## Models one person can realistically run

These fit a single good computer, or one graphics card in the larger cases. Sizes come from the Ollama library or Hugging Face.

| Model | Parameters | Practical size | Licence |
|---|---|---|---|
| [Gemma 4 E2B / E4B](https://huggingface.co/google/gemma-4-E4B-it) | 2 and 4 billion effective | About 5 to 10 GB | Apache 2.0 |
| [Gemma 4 12B](https://huggingface.co/google/gemma-4-12B-it) | 12 billion | About 8 GB | Apache 2.0 |
| [Nemotron 3.5 Bolt 74B-A9B](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF) | 74 billion, 9 active | 24 GB GGUF; recipes from 28 GB of memory | OpenMDW 1.1 |
| [gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) | 21 billion, 3.6 active | 14 GB | Apache 2.0 |
| [Muse Glimmer 30B](/en/learn/run-muse-glimmer-locally/) | About 30 billion | 16.8 GB (aim for 24 GB of memory) | Apache 2.0 |
| [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B) | 27 billion | 18 GB | Apache 2.0 |
| [Gemma 4 26B / 31B](https://huggingface.co/google/gemma-4-31B-it) | 26 billion (4 active), 31 billion | 16 to 19 GB for 26B | Apache 2.0 |
| [Nemotron 3.5 Lightning](/en/learn/nvidia-nemotron-3-5-models-compared/) | 30 billion, 3 active | About 22 GB NVFP4; one DGX Spark or H100 | OpenMDW 1.1 |

Note how friendly the licences are at this end. For most small projects, the question is quality on your tasks, not legal risk.

## Models for one server

These need data-centre GPUs or very large unified memory. Sizes are file totals on Hugging Face, in the most compact official format listed.

| Model | Total / active | Files | Hardware named by the maker | Licence |
|---|---|---|---|---|
| [gpt-oss-120b](https://huggingface.co/openai/gpt-oss-120b) | 117 billion / 5.1 billion | 65 GB (Ollama) | One 80 GB GPU | Apache 2.0 |
| [Kolibri-1](/en/learn/kolibri-1-aleph-alpha-sovereign-model/) | 78 billion / 3.46 billion | About 79 GB | One H200 or B200, or 2 H100s | Apache 2.0, weights only |
| [Nemotron 3.5 Super VL](/en/learn/nvidia-nemotron-3-5-models-compared/) | 120 billion / 12 billion | About 75 GB | One H200 or DGX Spark, or 2 H100s | OpenMDW 1.1 |
| [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) | About 180 billion | About 360 GB at BF16; a smaller FP8 version exists | Several GPUs at BF16 | Qwen Community 1.0 |
| [MiniMax M3.1-Flash](/en/learn/minimax-m3-open-weights-community-licence/) | 428 billion / 23 billion | About 252 GB | Several GPUs | MiniMax Community |
| [Beam](/en/learn/reflection-beam-open-weights-apache/) | 501 billion / 23 billion | About 294 GB | Several GPUs | Apache 2.0 |
| [GLM-5.3-Flash](/en/learn/glm-5-3-open-weights-licence/) | 320 billion / 18 billion | About 328 GB | Several GPUs | MIT |

Sparse designs show their value here. Beam's download is modest for 501 billion parameters because it comes in 4-bit, and few parameters are working on each token.

## Models that need a cluster, or a provider

| Model | Total / active | Files | What to expect | Licence |
|---|---|---|---|---|
| [DeepSeek-V4.1-Flash](/en/learn/deepseek-v4-1-flash-explained/) | 552 billion / 8 and 16 billion | About 510 GB | Multi-GPU server, 8-way example | MIT |
| [MiMo-V2.6-Pro](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL) | 1.02 trillion / 42 billion | About 574 GB | Multi-GPU server | MIT |
| [Mistral Large 4](/en/learn/mistral-large-4-open-weights-licence-hardware/) | 1 trillion / 49 billion | About 610 GB | 4 to 8 newest-generation GPUs | Modified MIT, service licence |
| [GLM-5.3](/en/learn/glm-5-3-open-weights-licence/) | 753 billion | About 756 GB | Multi-GPU server | GLM-5.3 License |
| [Inkling](https://huggingface.co/thinkingmachines/Inkling) | 975 billion / 41 billion | About 1.9 TB | Multi-GPU cluster | Apache 2.0 |
| [Muse Spark Open](/en/learn/muse-spark-open-weights-licence-mac-studio/) | 816 billion / 42 billion | 700 GB of memory at NVFP4 | Two 512 GB Mac Studios, or GPUs; form required | Muse Model License |
| [Kimi K3](/en/learn/kimi-k3-open-weights-licence-hardware/) | 2.8 trillion / 104 billion | About 1.56 TB | At least 8 of the latest GPUs | Kimi K3 License |
| [Qwen3.8 2.4T-A95B](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B) | 2.4 trillion / 95 billion | About 4.9 TB at BF16 | Multi-node | Qwen3.8-Max License |

Muse Spark Open stands out as the only one here with a real desktop route. Even so, two top-end Mac Studios are a serious purchase.

For almost everyone, these models are something you rent per token. But having the weights still has value: if a provider changes its terms or vanishes, someone else can host the same model.

## What you are legally allowed to do

| You want to | Apache 2.0, MIT, OpenMDW | Mistral Large 4 | Kimi K3 | GLM-5.3 | Qwen3.8-Max | Muse Spark Open | MiniMax Community |
|---|---|---|---|---|---|---|---|
| Use it internally | Yes | Yes | Yes | Yes | Yes | Yes, with usage policy | Non-commercial yes; business use needs a notice |
| Embed it in a paid product | Yes | Yes | Yes | Yes | Yes, with limits on work assistants | Yes, show "Built with" | Yes, show "Built with" and send notice |
| Fine-tune and share it | Yes | Yes | Yes | Yes | Yes | Yes, derivative rules | Yes, on the same conditions |
| Sell API access | Yes | Always needs a licence | Agreement above $20M US | Security review above $10B US | Licence above $50M US | Agreement above $20M US | Authorization above $20M US |
| Use for military purposes | Licence is silent | Licence is silent | Licence is silent | Licence is silent | Licence is silent | Check usage policy | Banned |

This table summarizes the licence files as of this date. It isn't legal advice, and licences can change with each model, as they did for GLM and Kimi. Copyright, privacy and trade laws still apply whatever the licence says. In Canada, a business handling personal information still has its privacy duties, including under Québec's Law 25, wherever the model runs.

> **Try it:** Write down your machine's graphics memory (or a Mac's unified memory), then, as a rule of thumb, pick the largest model above whose file is at least 25 percent smaller than that figure. Run it through Ollama or LM Studio and give it one task you do every week.

## How do you choose in ten minutes?

**Minute 1: where can your data go?** If it may go to a provider, hosted access opens up every model above. If it must stay on your machines, your hardware sets the limit.

**Minutes 2 to 3: what memory do you have?** Under 16 GB, begin with Gemma 4 E2B or E4B. At 24 to 32 GB, try Qwen3.8-27B, Muse Glimmer or gpt-oss-20b. With one 80 GB data-centre GPU, gpt-oss-120b fits, and at 160 GB or more you can run Kolibri-1 or Super VL. Beyond that, plan a server.

**Minutes 4 to 6: what will you do with it?** For your own work, or a product with the model built in, most licences are fine. If you plan to sell access to the model itself, cross off Mistral Large 4 unless you are ready to talk to Mistral, and check thresholds on the others.

**Minutes 7 to 8: do you need something specific?** For French, test Mistral's models and Gemma's multilingual support. For German, try Kolibri. For audio, look at Gemma 4's smaller sizes or MiMo. For image input, avoid Beam, which handles only text. For full transparency about training data, look at Nemotron or Switzerland's public Apertus.

**Minutes 9 to 10: test.** Benchmark tables are mostly self-reported, and each lab picks the setup that suits it. Run five real tasks on your two best candidates through a host, score them, and only then buy hardware.

## Sources

- gpt-oss-120b model card: https://huggingface.co/openai/gpt-oss-120b
- Gemma 4 31B model card: https://huggingface.co/google/gemma-4-31B-it
- qwen3.8 - Ollama library: https://ollama.com/library/qwen3.8

---

From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/which-open-weight-model-compared/, no commercial use without permission, and keep the same licence.