Open-Weight LLMs Ranked by License: What You Can Actually Ship
As of 9 November 2026, Gemma 4 31B and Qwen3.8-27B are the easiest open-weight models to ship: Apache 2.0, cheap to rent, workstation-sized. Nemotron 3.5 Super VL and DeepSeek V4.1-Flash-1017 follow. Mistral Large 4, Kimi K3.1 Pro, Muse Spark Open and MiniMax M3.1-Flash need separate agreements, attribution or notices once you resell access or pass revenue thresholds.
Gemma 4 31B and Qwen3.8-27B are the easiest open-source LLMs to ship in a commercial product as of 9 November 2026: both use the Apache 2.0 license, cost little to rent and fit on workstation-class hardware. At the other end, Muse Spark Open and Kimi K3.1 Pro attach attribution duties and revenue thresholds that matter the moment you resell access. Strictly, almost none of these is open source in the Open Source Initiative's sense, which also requires training-data information and training code; they are open-weight models, meaning downloadable weights under a license.
Lists of the best open source LLMs usually rank by benchmark scores, and so did our May roundup of open-source LLMs. This guide ranks ten current releases by what their licenses let you do, then by hosted price and self-hosting footprint. Every license term was read from the license file itself on 9 November 2026, and every version was checked against the vendor's Hugging Face organization that day.
How these open-source LLMs are scored
Each model gets three scores out of 10, combined with the same fixed weights for every entry:
- License terms (50%): start at 10. Subtract 1 for a vendor-specific license rather than Apache 2.0 or MIT, and 2, 3 or 4 when API resale needs a separate agreement or review (above US$1 billion in revenue, at US$50 million or less, or at any revenue). Subtract 1 for a name display required only at scale or 2 if required of everyone, 1 for an attached usage policy, and 2 where the base grant is non-commercial.
- Hosted price (25%): a 3:1 input-to-output blend of the price per million tokens on the named provider's pricing page on 9 November 2026. Blends up to $0.30 score 10, $0.60 score 8, $1.20 score 6, $2.50 score 4 and $5 score 2; no per-token price scores 0.
- Self-hosting footprint (25%): a hardware class banded on the vendor's stated total parameters: under 40B workstation (10), 40B-200B multi-GPU server (7), 200B-500B large server (5), 500B-1.1T data-center node (3), above 1.1T multi-node (1). The bands are our classification, not vendor requirements; vendor hardware statements are quoted where they exist.
Ties go to the higher license score. Vendor-reported benchmarks play no part in the scores.
| Rank | Model (version, date) | License | Price | Footprint | Total |
|---|---|---|---|---|---|
| --- | --- | --- | --- | --- | --- |
| 1 | Gemma 4 31B (2 Apr 2026) | 10 | 10 | 10 | 10.00 |
| 2 | Qwen3.8-27B (repo 5 Aug 2026) | 10 | 8 | 10 | 9.50 |
| 3 | Nemotron 3.5 Super VL 120B-A12B (15 Oct 2026) | 9 | 10 | 7 | 8.75 |
| 4 | DeepSeek V4.1-Flash-1017 (repo 17 Oct 2026) | 10 | 8 | 3 | 7.75 |
| 5 | K-EXAONE 2.0 750B-A37B (repo 29 Jul 2026) | 10 | 0 | 3 | 5.75 |
| 6 | GLM-5.3 (repo 25 Aug 2026) | 7 | 4 | 3 | 5.25 |
| 7 | Mistral Large 4 (API 6 Oct 2026) | 5 | 6 | 3 | 4.75 |
| 8 | MiniMax M3.1-Flash (repo 17 Oct 2026) | 1 | 10 | 5 | 4.25 |
| 9 | Kimi K3.1 Pro (repo 2 Nov 2026) | 5 | 2 | 1 | 3.25 |
| 10 | Muse Spark Open 816B-A42B (15 Oct 2026) | 3 | 4 | 3 | 3.25 |
Open-source LLMs vs open-weight models: what is the difference?
An open-weight model publishes its trained parameters under a license; an open-source AI system, under the OSI's Open Source AI Definition 1.0, must also provide detailed training-data information and the complete training and data-processing code under OSI-approved terms. Most open-source LLMs people search for meet only the first test, so "open-weight" is the accurate term.
An Apache 2.0 or MIT license is OSI-approved, but that covers the license text, not the missing data and code. The AI2 sidebar below covers the one model here that publishes all three.
Quick license answers for new releases
Three recent releases draw the most license questions. Each answer below was checked against the license file in the vendor's Hugging Face repository on 14 November 2026, and the ranked entries further down give the full terms.
Can I use Kimi K3.1 commercially?
Yes, with conditions. As of 14 November 2026, the Kimi K3.1-Pro License allows commercial use, but a Model-as-a-Service business above US$20 million in 12-month revenue needs a separate Moonshot AI agreement. Products above 100 million monthly active users, or US$20 million in monthly revenue, must display "Kimi K3.1-Pro". Internal use is exempt. The license file defines Model-as-a-Service as giving third parties inference or fine-tuning access, and excludes products where the model is only embedded in specific features.
Is Qwen 3.8 Apache 2.0?
Only the 27B checkpoint. The Qwen3.8-27B repository carries the Apache 2.0 license, as read on 14 November 2026. The 2.4T-A95B flagship uses the Qwen3.8-Max License, and Qwen3.8-Flash-Next uses Qwen Community License 1.0. Both custom licenses require a separate license for Model-as-a-Service or AI Work Assistant businesses, above US$50 million in revenue for the flagship. The licenses are in the Qwen3.8-27B, Qwen3.8-2.4T-A95B and Qwen3.8-Flash-Next repositories.
Is DeepSeek V4.1 MIT licensed?
Yes. Both DeepSeek V4.1 repositories on Hugging Face, DeepSeek-V4.1-Flash-1017 and DeepSeek-V4.1-Flash, carry the MIT License, as read on 14 November 2026. The only condition is keeping the copyright and permission notice. This covers V4.1 only; DeepSeek publishes other V4 checkpoints in separate repositories, so check the license file of each one you download. Read the DeepSeek-V4.1-Flash-1017 and DeepSeek-V4.1-Flash license files.
Open-source LLM licenses compared: thresholds, attribution and usage policies
| Model | License | Threshold that triggers extra terms | Attribution duty | Usage policy |
|---|---|---|---|---|
| --- | --- | --- | --- | --- |
| Gemma 4 31B | Apache 2.0 | None | Keep license and notices when redistributing | None in the license |
| Qwen3.8-27B | Apache 2.0 | None | Keep license and notices when redistributing | None in the license |
| Nemotron 3.5 Super VL | OpenMDW-1.1 | None (rights end if you sue claiming the model infringes) | Keep agreement and notices when distributing | None; no restrictions on outputs |
| DeepSeek V4.1-Flash-1017 | MIT | None | Keep copyright and permission notice | None |
| K-EXAONE 2.0 | Apache 2.0 | None | Keep license and NOTICE file | None in the license |
| GLM-5.3 | GLM-5.3 License | API resale with group revenue above US$10bn in 12 months: Z.AI security review | Keep copyright notice | Comply with applicable law |
| Mistral Large 4 | Mistral Ecosystem Open-Weight License | Any Model-as-a-Service business: separate commercial license | Keep copyright notice | None |
| MiniMax M3.1-Flash | MiniMax Community License | Any commercial use: notice; above US$20m yearly revenue: prior written authorization | "Built with MiniMax M3.1-Flash" for any commercial use | Appendix of prohibited uses, including military |
| Kimi K3.1 Pro | Kimi K3.1-Pro License | API resale above US$20m in 12 months: separate agreement | "Kimi K3.1-Pro" in the UI above 100m MAU or US$20m monthly revenue | Comply with applicable law |
| Muse Spark Open | Muse Model License | API resale above US$20m gross revenue in 12 months: notify Meta, written agreement | "Built with Muse Spark Open" for every product using it | Exhibit A usage policy; not intended for under-18s |
Most custom open-source LLM licenses target one activity, reselling access. Three of them (GLM, Kimi and Mistral) define that Model-as-a-Service activity almost identically: giving a third party inference or fine-tuning access "in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data." All three exclude end-user products where model capabilities are "solely embedded within specific features or harnesses."
The ten open-source LLMs, ranked by license
1. Gemma 4 31B: Apache 2.0 with nothing added
2. Qwen3.8-27B: clean license, split family
3. Nvidia Nemotron 3.5 Super VL 120B-A12B: permissive, and Nvidia names the hardware
4. DeepSeek V4.1-Flash-1017: MIT at a low price, but large
5. LG K-EXAONE 2.0 750B-A37B: Apache 2.0 at 750B scale
6. GLM-5.3: one condition, aimed at giants
7. Mistral Large 4: fine to embed, not to resell
8. MiniMax M3.1-Flash: cheap to rent, strict to use
9. Kimi K3.1 Pro: the largest model, the priciest to rent
10. Muse Spark Open 816B-A42B: attribution for everyone
Sidebar: the most open option is AI2's OLMo 3.1 (December 2025)
The Allen Institute for AI's Olmo 3.1 32B Instruct, published on 10 December 2025, is older than every ranked entry, so it sits outside the ranking. It is listed for one reason: AI2 publishes the weights, the Dolma 3 pre-training data (on Hugging Face under ODC-BY), the Dolci post-training datasets and its training and evaluation code. Among open-source LLMs, that comes closest to what the OSI definition asks for.
The card states Apache 2.0, though the repository has no LICENSE file of its own, and says the model "is intended for research and educational use" under AI2's Responsible Use Guidelines. AI2's newest language model is the smaller Olmo-Hybrid 7B (January-February 2026). Choose OLMo when you need to inspect the training data itself, not for raw capability.
What can you actually ship with each license?
Internal tools are safe on all ten, with care on two. Product features that embed a model need no extra steps on eight at ordinary scale. Reselling API access is where licenses diverge: five models allow it outright, GLM-5.3 only reviews companies above US$10 billion in revenue, and Mistral, Kimi, Muse and MiniMax require agreements, notices or authorization.
An internal tool
Every license here allows it. Kimi, Mistral and the Qwen custom licenses exempt internal use outright, and Meta's revenue clause excludes it.
Two cautions apply. Muse Spark Open's display duty covers "each product or service" with no internal carve-out, and MiniMax's license has no internal-use exemption at all, so commercial internal use calls for its notice. If you plan to run models on your own machines, our guide to running LLMs locally covers the tooling, and the self-hosted coding agent walkthrough shows one internal setup end to end.
A SaaS feature
Embedding a model inside a feature, such as summarizing tickets in a help desk, is not Model-as-a-Service under the GLM, Kimi and Mistral definitions. Meta's revenue clause excludes powering "features of the Licensee's own product". The duties that remain are display rules: "Built with Muse Spark Open" always, "Built with MiniMax M3.1-Flash" plus a notice always, and "Kimi K3.1-Pro" above 100 million monthly users. Whatever you ship, trace it; our comparison of AI observability platforms covers the tools for that.
Reselling API access
Five of these open-source LLMs add no terms for resale: Gemma 4, Qwen3.8-27B, Nemotron 3.5 Super VL, DeepSeek V4.1-Flash-1017 and K-EXAONE 2.0. GLM-5.3 asks for a security review only above US$10 billion.
Kimi K3.1 Pro and Muse Spark Open need an agreement above US$20 million. Mistral Large 4 needs a commercial license from the first customer, and MiniMax needs a notice or authorization depending on revenue. Our comparison of LLM inference providers shows how the hosts that resell these models price them.
Before you ship an open-source LLM
- Read the LICENSE file in the exact repository you download from: open-source LLM licenses can differ between variants of one family.
- Check the Hugging Face organization for a newer checkpoint, as DeepSeek's -1017 and MiniMax's M3.1-Flash both arrived on 17 October.
- Count your affiliates' revenue as well as your own: the GLM, Kimi, Muse and Qwen thresholds aggregate them.
- Re-check hosted prices on the day you budget; DeepInfra showed promotional discounts on DeepSeek and GLM models on 9 November 2026.
Sources
Pages and license files read on 9 November 2026 (the Kimi, Qwen and DeepSeek license files were read again on 14 November 2026):
- Gemma 4 model card (Hugging Face) (huggingface.co)
- Gemma 4: Expanding the Gemmaverse with Apache 2.0 (Google Open Source Blog) (opensource.googleblog.com)
- Gemma 4 license (Google AI for Developers) (ai.google.dev)
- Qwen3.8-27B LICENSE (huggingface.co)
- Qwen3.8-Max License (Qwen3.8-2.4T-A95B) (huggingface.co)
- Qwen Community License 1.0 (Qwen3.8-Flash-Next) (huggingface.co)
- OpenMDW License Agreement, version 1.1 (openmdw.ai)
- DeepSeek-V4.1-Flash-1017 LICENSE (huggingface.co)
- DeepSeek-V4.1-Flash LICENSE (huggingface.co)
- DeepSeek API models and pricing (api-docs.deepseek.com)
- K-EXAONE 2.0 LICENSE (huggingface.co)
- K-EXAONE 2.0 on FriendliAI (friendli.ai)
- GLM-5.3 LICENSE (huggingface.co)
- Z.ai pricing (docs.z.ai)
- Mistral Large 4 LICENSE (huggingface.co)
- Mistral Large 4 model docs and pricing (Mistral) (docs.mistral.ai)
- MiniMax-M3.1-Flash LICENSE (huggingface.co)
- MiniMax API pricing (platform.minimax.io)
- Kimi K3.1 Pro LICENSE (huggingface.co)
- Kimi API model inference pricing (platform.kimi.ai)
- Muse Model License, Muse Spark Open version (LICENSE.md, gated) (huggingface.co)
- Together AI pricing (together.ai)
- DeepInfra pricing (deepinfra.com)
- Qwen3.8-27B endpoints and prices (OpenRouter) (openrouter.ai)
- Olmo 3.1 32B Instruct model card (AI2) (huggingface.co)
- Dolma 3 pre-training data pool (AI2, Hugging Face) (huggingface.co)
- The Open Source AI Definition 1.0 (Open Source Initiative) (opensource.org)
Key Takeaways
- Four of the ten models (Gemma 4 31B, K-EXAONE 2.0, Qwen3.8-27B and DeepSeek V4.1-Flash-1017) use Apache 2.0 or MIT with no added conditions, as of 9 November 2026.
- Mistral Large 4's license requires a separate commercial license from anyone running a Model-as-a-Service business, with no revenue floor.
- Muse Spark Open requires every product that uses it to display "Built with Muse Spark Open", and its usage policy says the model is not intended for anyone under 18.
- MiniMax M3.1-Flash is licensed for non-commercial use by default: commercial users must send MiniMax a notice, or get prior written authorization above US$20 million in yearly revenue.
- Kimi K3.1 Pro needs a separate Moonshot AI agreement for API resellers above US$20 million in 12-month revenue, and a "Kimi K3.1-Pro" display above 100 million monthly users.
- The same family can carry different licenses: Qwen3.8-27B is Apache 2.0, but the 2.4T flagship and Qwen3.8-Flash-Next use custom Qwen licenses.
- Only Meta and Nvidia state hardware requirements for their entries; Meta says Muse Spark Open needs at least 700GB of VRAM at NVFP4, and Nvidia lists four H100 GPUs for Nemotron 3.5 Super VL.
Frequently Asked Questions
What are the best open source LLMs for commercial use?
Ranked by license terms, hosted price and self-hosting footprint as of 9 November 2026, Gemma 4 31B and Qwen3.8-27B lead: both are Apache 2.0 and fit on workstation-class hardware. Nemotron 3.5 Super VL (OpenMDW-1.1) and DeepSeek V4.1-Flash-1017 (MIT) follow. All four can be resold as an API without extra agreements.
Is an open-weight model the same as an open-source model?
No. Open-weight means you can download the trained parameters under a license. The Open Source Initiative's Open Source AI Definition also requires detailed training-data information and the complete training code under approved terms. Most popular models publish weights and inference code only, so open-weight is the accurate label for them.
Can I sell API access to an open-weight model?
It depends on the license. Apache 2.0, MIT and OpenMDW-1.1 models carry no extra terms for resale. Mistral Large 4 needs a commercial license for any Model-as-a-Service business. Kimi K3.1 Pro and Muse Spark Open need a separate agreement above US$20 million in 12-month revenue, and GLM-5.3 a security review above US$10 billion.
Do I have to credit the model in my product?
Sometimes. Muse Spark Open requires "Built with Muse Spark Open" for every product using it, and MiniMax M3.1-Flash requires "Built with MiniMax M3.1-Flash" for any commercial use. Kimi K3.1 Pro and the custom Qwen licenses require the model name only above 100 million monthly users or US$20 million monthly revenue.
Which open-weight LLM can I run on a single workstation?
Of the ten models here, Gemma 4 31B and Qwen3.8-27B are the only ones in the workstation class. Google's model card places the 31B among models for consumer GPUs and workstations. Nvidia's Nemotron 3.5 Super VL is the next step up: Nvidia lists four H100 or two GB200 GPUs for its BF16 version.
Does using an internal tool trigger these license conditions?
Usually not. The Kimi, Mistral and Qwen licenses exempt internal use outright, and Meta's revenue clause excludes it too. Two exceptions need care: Muse Spark Open's attribution duty has no internal-use carve-out, and MiniMax's license has no internal-use exemption, so commercial internal use still calls for its notice.
About the Author
Aisha Patel
AI Editorial Desk
AI Editorial Desk · Web3AIBlog
Aisha Patel is a pen name for our AI editorial desk. Posts under this byline are written and reviewed by our team of contributors with backgrounds in machine learning, large language models, AI infrastructure, and applied research. The desk covers frontier model releases, agent architectures, retrieval-augmented generation, on-device inference, and the engineering tradeoffs that matter when shipping AI in production. Every technical claim is verified against primary sources before publication.