Open Model Licenses | How Much To Run AI
Reference the licenses, access conditions, and practical restrictions for the open-weight models in the current calculator catalog.
Last updated: 2026-09-21
This page records the license information for the model entries currently in the calculator catalog. Download access alone does not establish that a model is open source or unrestricted for commercial use. Read the license file and acceptable-use terms for the exact repository and checkpoint you plan to deploy.
License summaries are for planning and comparison only. They are not legal advice. The repository's license, model card, access terms, and applicable law control.
How to read this page
- Open weights means that the publisher makes model weights available under stated access terms. It does not by itself mean that the training data, code, or every model component is open source.
- Open source is a separate question that depends on the license and the definition being used. This page does not apply an OSI certification to any model.
- Commercial use needs more than a yes/no label. Restrictions may cover attribution, notices, trademarks, acceptable use, distribution, scale, monthly active users (MAU), model-as-a-service (MaaS), or additional agreements.
- The calculator's model records now include repository, variant, context window, weight format, license name, license URL, gated-access flag, and the date the metadata was last manually checked. These fields are still maintained by hand and are not automatically synchronized with the repository.
Current catalog
| Calculator entry | Variant | Repository / checkpoint | License | Gated | Practical restrictions and notes |
|---|---|---|---|---|---|
| Qwen 3.8 27B | Post-trained vision-language | Qwen/Qwen3.8-27B | Apache License 2.0 | No | This dense multimodal checkpoint includes a vision encoder and supports flexible thinking control. Qwen publishes BF16 Safetensors; Ollama exposes a Q4_K_M artifact as qwen3.8:27b. Preserve the required Apache-2.0 notices and verify the exact quantized artifact before redistribution. |
| Qwen 3.8 Flash-Next | Post-trained vision-language MoE | Qwen/Qwen3.8-Flash-Next | Qwen Community License 1.0 (custom) | No | Qwen4 architecture preview: 125B main parameters + 51B n-gram embeddings + 4B MTP, ~6B activated per token; official BF16 and FP8 Safetensors. Use and redistribution are broadly permitted with notices preserved, but products above 100M MAU or USD 20M monthly revenue must prominently display the model name, and MaaS or AI work-assistant businesses need a separate Qwen license before commercial use (internal use excepted). |
| Gemma 4 E2B | Instruction-tuned multimodal | google/gemma-4-E2B-it | Apache License 2.0 | No | Google reports 2.3B effective parameters and 5.1B including embeddings. The checkpoint supports text, image, and audio input with 128K context. Its model-card metadata links to Google's Gemma 4 Apache-2.0 license; Ollama provides the Q4_K_M tag gemma4:e2b-it-q4_K_M. |
| Gemma 4 E4B | Instruction-tuned multimodal | google/gemma-4-E4B-it | Apache License 2.0 | No | Google reports 4.5B effective parameters and 8B including embeddings. The checkpoint supports text, image, and audio input with 128K context; Ollama's exact Q4_K_M tag is gemma4:e4b-it-q4_K_M. |
| Gemma 4 12B Unified | Instruction-tuned unified | google/gemma-4-12B-it | Apache License 2.0 | No | The encoder-free unified checkpoint directly projects text, image, and audio inputs into a decoder-only model and supports 256K context. Ollama's exact Q4_K_M tag is gemma4:12b-it-q4_K_M. |
| Gemma 4 26B A4B | Instruction-tuned MoE | google/gemma-4-26B-A4B-it | Apache License 2.0 | No | This text-and-image MoE checkpoint has about 25.8B complete checkpoint parameters and 3.8B active language-model parameters. Capacity estimates use the complete weights, not A4B. Ollama's Q4_K_M tag is gemma4:26b-a4b-it-q4_K_M. |
| Gemma 4 31B | Instruction-tuned dense | google/gemma-4-31B-it | Apache License 2.0 | No | This dense text-and-image checkpoint has about 31.3B complete checkpoint parameters and 256K context. Ollama's exact Q4_K_M tag is gemma4:31b-it-q4_K_M. |
| Llama 3.1 8B | Instruct | meta-llama/Llama-3.1-8B-Instruct | Llama 3.1 Community License | Yes | Access is gated and requires accepting Meta's terms. Commercial use is subject to the Community License and Acceptable Use Policy. The license includes a separate-authorization threshold for products with more than 700 million monthly active users and naming/attribution requirements for derivatives. |
| Llama 3.1 70B | Instruct | meta-llama/Llama-3.1-70B-Instruct | Llama 3.1 Community License | Yes | The same Community License, gated access, acceptable-use, scale threshold, and derivative naming/attribution conditions apply. Check the current Meta terms before a large-scale launch. |
| DeepSeek-V3 | Base | deepseek-ai/DeepSeek-V3 | MIT | No | MIT generally permits commercial use, modification, and redistribution when the copyright and license notice are retained. Verify the model card and any bundled component terms; the calculator entry is a 671B total / 37B active MoE record. |
| DeepSeek-V4-Flash-0731 | GA (0731) | deepseek-ai/DeepSeek-V4-Flash-0731 | MIT | No | The official model card identifies this as the official release superseding Preview. It keeps the same model size and structure while adding the DSpark speculative-decoding module. The catalog entry describes a 284B total / 13B active MoE model and native FP4+FP8 weights; the hosted API announcement calls the release public beta. |
| DeepSeek-V4.1-Flash | GA (2026-09-10) | deepseek-ai/DeepSeek-V4.1-Flash | MIT | No | This multimodal MoE release has about 749B total parameters for capacity estimates: 552B backbone plus about 196B Engram. It activates 8B during prefill and 16B during decode, supports 1M context with YaRN from 65,536, and uses an MXFP4/MXFP8 mixed Safetensors checkpoint. The official hosted API name is deepseek-flash. |
| DeepSeek-V4-Pro-0813 | GA (0813) | deepseek-ai/DeepSeek-V4-Pro-0813 | MIT | No | The official GA announcement and the model card identify this checkpoint as superseding Preview. The official checkpoint uses Safetensors with FP4 expert weights and FP8 mixed quantization; it does not restate a separate active-parameter count. The calculator uses about 1.7T total parameters for capacity estimates, not a performance or production guarantee. |
| GLM-5.2 | Instruct | zai-org/GLM-5.2 | MIT (model-card reference) | No | MIT generally permits commercial use, modification, and redistribution with the required notices. Verify the exact model card and official release because published parameter counts and checkpoint metadata may differ across sources. |
| GLM-5.3 | Post-trained | zai-org/GLM-5.3 | GLM-5.3 License (custom) | No | This open-weight checkpoint is released under a custom license. It permits use, modification, deployment, and redistribution with required notices, but a Model as a Service business whose aggregate revenue exceeds USD 10 billion over any consecutive 12 months must pass Z.AI's security review before commercial use. |
| GLM-5.3-Flash | Post-trained multimodal | zai-org/GLM-5.3-Flash | MIT | No | This natively multimodal MoE checkpoint has 320B total and 18B active parameters, with FP8/BF16 Safetensors and a 1M context window. Preserve the MIT notice and verify any separately licensed quantized artifact before redistribution. |
| Kimi K3 | Instruct | moonshotai/Kimi-K3 | Kimi K3 License (custom) | No | This is not an unconditional MIT or Apache license. The published terms require a separate agreement for MaaS businesses above USD 20 million in annual revenue, and require products at or above 100 million MAU or USD 20 million in monthly revenue to prominently display “Kimi K3”. Read the current license before offering the model as a service or redistributing it. |
For non-developers
The table above keeps the license names and source summaries. The notes below explain what those entries mean in everyday terms. They are a reading aid, not new license terms.
- Qwen 3.8 27B: The exact checkpoint is
Qwen/Qwen3.8-27B, and its default Ollama tag isqwen3.8:27b. It uses Apache-2.0; check the notices and terms attached to the exact multimodal or quantized artifact you redistribute. - Qwen 3.8 Flash-Next: The exact checkpoint is
Qwen/Qwen3.8-Flash-Next(an official FP8 variant lives in the same repository with a-FP8suffix); no Ollama tag exists yet. It uses the custom Qwen Community License 1.0, not Apache-2.0: large products must prominently display the model name, and MaaS or AI work-assistant businesses need a separate Qwen license first. - Gemma 4: The calculator uses the exact instruction-tuned E2B, E4B, 12B Unified, 26B A4B, and 31B checkpoints. Google publishes Gemma 4 under Apache-2.0 and the repositories are not gated. Preserve required notices and verify the exact checkpoint and quantized artifact before redistribution.
- Llama 3.1 8B and 70B: You must pass Meta's gated-access process and accept its terms. Commercial use also follows the Community License and Acceptable Use Policy. Large products and derivative models may have extra requirements.
- DeepSeek-V3: The entry points to MIT terms, so keeping the copyright and license notice is the main baseline condition in the summary. Check the model card and bundled components before distribution.
- DeepSeek-V4-Flash-0731: The official model card identifies this MIT release as superseding Preview. It keeps the same model size and structure and adds DSpark speculative decoding. The listed FP4+FP8 format and very large model size need a separate deployment check; DeepSeek-hosted API access is separately documented as public beta.
- DeepSeek-V4.1-Flash: The exact checkpoint is published under MIT and is natively multimodal. The catalog records about 749B total parameters from the model card's 552B backbone plus about 196B Engram, with 8B prefill and 16B decode active parameters. The official hosted API name is
deepseek-flash; preserve the MIT notice and verify the exact checkpoint and bundled component terms before redistribution. - DeepSeek-V4-Pro-0813: The official model card states that the repository and weights use MIT and that this release supersedes Preview. Preserve the license notice. Its FP4-expert plus FP8-mixed representation is not a generic INT4 download, and the calculator's approximately 1.7T total-parameter figure is a capacity-estimation input rather than a performance guarantee. DeepSeek-hosted services have separate terms.
- GLM-5.2: The entry refers to MIT terms from the model card. Verify the current model card and release details for the exact checkpoint before use.
- GLM-5.3: This is a custom Z.AI license, not MIT. It includes a security-review condition for very large Model as a Service businesses; read the exact license before commercial hosting.
- GLM-5.3-Flash: This checkpoint uses MIT and is natively multimodal. Preserve the notice and check the exact checkpoint and any quantized artifact before redistribution.
- Kimi K3: This is a custom license with conditions for some MaaS and very large commercial products. Read the current terms before offering it as a service or redistributing it.
Practical checklist before commercial deployment
For the exact checkpoint, confirm:
- The repository, revision, and whether it is Base, Instruct, Chat, Preview, or another release state.
- The license file, model card, acceptable-use policy, and any gated-access terms.
- Attribution, copyright, notice, trademark, naming, and redistribution requirements.
- Any MAU, revenue, MaaS, geography, user, or use-case thresholds.
- Whether quantized files, adapters, tokenizer files, inference code, and other bundled components have separate terms.
- Whether the planned deployment is direct inference, an internal tool, an API, or a model-as-a-service product.
The cost calculator can help estimate infrastructure spending, but it does not grant permission to download, modify, host, or commercialize a model.
Sources
The repository links in the table are the primary starting points. License terms can change independently of calculator pricing data, so check the repository's current LICENSE, model card, and publisher terms on the date of deployment. Metadata in the calculator is maintained manually and stamped with a checked-at date; use it as a reference, not as a real-time compliance check.
Image model: Qwen-Image-2.1
Verified September 21, 2026. The Qwen Research License Agreement limits use to non-commercial research or evaluation. Commercial use requires a separate license from the publisher. Downloadable weights do not imply unrestricted commercial use. The image cost calculator excludes commercial license fees.