An open-weight model is a large language model whose trained parameters (weights) the publisher lets anyone download and run, while a closed model is one you can only reach through the publisher's own API, apps or approved cloud partners. The difference decides where the model runs, who sees your prompts, what the licence lets you do and how you pay. This page explains the terms in plain language, shows how current models from named providers fit each category, and sets out the factual criteria teams use when choosing between them. It is written for anyone evaluating models, from developers to procurement and governance teams.
Last checked: 2026-11-15. Providers are listed alphabetically; order is never a ranking.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.
Key takeaways
- Open weights are not the same as open source. The Open Source Initiative's Open Source AI Definition 1.0 also requires training code and detailed data information, which most open-weight releases do not publish in full.
- The licence is the real dividing line. Open-weight releases range from Apache 2.0 to custom licences written by the publisher, so "downloadable" says nothing about commercial or redistribution rights until you read the licence file.
- Closed models are reached only through an API or app. The provider runs the model, sets the price per token and decides retention, access tiers and retirement dates.
- Many providers publish both kinds. Google (Gemini and Gemma) and OpenAI (GPT and gpt-oss) each document closed and open-weight families side by side.
- Running open weights moves cost and risk to you. Hardware, serving software, patching, model file provenance and access control become your responsibility rather than the provider's.
What does open-weight mean?
A model's weights are the billions of numbers learned during training. When a publisher releases them, usually on Hugging Face, you can load the model on your own hardware with serving software such as Transformers, vLLM, SGLang or TokenSpeed, all of which the Qwen3.8-27B model card names as compatible. You can also fine-tune the weights on your own data, quantise them to fit smaller GPUs, or run them fully offline.
What you usually do not get is the training data, the full training code or the ability to reproduce the model. That is why the industry uses "open-weight" rather than "open source" for most releases.
What is a closed model?
A closed (or proprietary) model never leaves the provider's infrastructure. You send prompts to an endpoint and receive outputs, billed per token. The provider controls the model version, the context window offered, rate limits, data retention and the date a model is retired. Some closed models are also restricted to approved customers: Anthropic's documentation states that Claude Mythos 5.2 is available only to organisations verified through Anthropic's verification programmes, such as its Cyber Verification Program. Check each provider's model page for access status before planning around a model.
Is an open-weight model open source?
Usually not in the strict sense. The Open Source Initiative published the Open Source AI Definition (OSAID) version 1.0 in October 2024. It requires four freedoms: to use the system for any purpose, study how it works, modify it, and share it. As a precondition it requires the "preferred form to make modifications", which the definition says must include all of these elements:
| OSAID element | What it requires | Typical open-weight release |
|---|---|---|
| Parameters | The model weights and configuration settings | Published |
| Code | The complete source code used to train and run the system | Inference code usually published; full training code varies by release |
| Data information | Sufficiently detailed information about the training data for a skilled person to build a substantially equivalent system | Varies by release; often a summary only |
Some releases go further than weights. NVIDIA's model card for Nemotron 3 Ultra 550B-A55B states that NVIDIA released its final pre-training and post-training data alongside the model (with some subsets gated behind approval) and that the end-to-end training recipe is in the NVIDIA Nemotron Developer Repository. Whether any specific release meets OSAID is a judgement for the OSI and the reader, not something this page assesses.
Open-weight vs closed: how do they compare?
As of 2026-11-15, from provider documentation.
| Factor | Open-weight model | Closed model |
|---|---|---|
| Where it runs | Your hardware, your cloud account, or any host that serves it | Provider API, provider apps, or cloud partners the provider authorises |
| Access to weights | Download, usually from Hugging Face or Kaggle | None |
| Licence | Set by the publisher: Apache 2.0, MIT or a custom licence | Provider terms of service and commercial terms |
| Cost model | Hardware, energy and operations; or per-token fees if you use a hosted endpoint | Per-token fees, plus tool and storage charges |
| Fine-tuning | Full control over method and data | Only where the provider offers a fine-tuning service |
| Versioning | The file you downloaded does not change | Provider updates, deprecates and retires model IDs on its own schedule |
| Data handling | Prompts stay wherever you run the model | Governed by the provider's retention and training-use terms |
| Safety controls | Whatever you add; the publisher's safeguards can be changed by anyone who fine-tunes | Provider-side classifiers, usage policies and access tiers |
Which current models are open-weight and which are closed?
As of 2026-11-15. Providers alphabetical. Each row cites the official page fetched on that date. "Not stated by <provider>" means the page did not give a value.
| Provider | Model | Open weights | Licence | Parameters (total / active) | Context window | Source |
|---|---|---|---|---|---|---|
| Alibaba (Qwen) | Qwen3.8-27B | Yes | Apache 2.0 | 27B (dense) | 262,144 native, up to 1,000,000 | Hugging Face model card |
| Anthropic | Claude Fable 5.2 | No | Proprietary | Not stated by Anthropic | 1M | Anthropic Claude Fable 5.2 model page |
| DeepSeek | DeepSeek-V4.1-Flash-1017 | Yes | Not stated on the DeepSeek Hugging Face organisation page | 763B (Hugging Face size) | 1M (API) | Hugging Face organisation page; DeepSeek pricing page |
| Gemma 4 | Yes | "Gemma 4 license" (linked to an Apache 2.0 page) plus Gemma terms of use | Not stated by Google on the docs page | 128K, up to 256K | Google Gemma docs | |
| Meta | Muse Spark Open | Yes | Not stated by Meta on the page fetched | 816B | Not stated by Meta | dev.meta.ai |
| Mistral AI | Mistral Large 4 | Yes | Mistral Ecosystem Open Weight License | 1.05T / 52B | 1M | Mistral model card |
| Mistral AI | Mistral Small 4 | Yes | Apache 2.0 | 119B / 6.5B | 256K | Mistral model card |
| Moonshot AI | Kimi K3.1 Pro | Yes | Kimi K3.1 Pro License | 2.8T / 104B | 1,048,576 | Hugging Face model card |
| NVIDIA | Nemotron 3 Ultra 550B-A55B | Yes | OpenMDW-1.1 | 550B / 55B | Up to 1M | Hugging Face model card |
| OpenAI | gpt-oss-120b | Yes | Apache 2.0 | 117B / 5.1B | 131,072 | OpenAI model page |
| OpenAI | GPT-6.1 Sol | No | Proprietary | Not stated by OpenAI | 1.05M | OpenAI GPT-6.1 Sol model page |
| xAI | Grok 4.7 | No | Proprietary | Not stated by xAI | 500K | xAI pricing page |
Several providers publish both. Google documents Gemma as "open models built from the same research and technology used to create the Gemini models", while Gemini itself is API-only. OpenAI's gpt-oss-120b page describes it as its "most powerful open-weight model", alongside its closed GPT family. Mistral AI publishes open models under Apache 2.0 (Mistral Small 4) and under its own Mistral Ecosystem Open Weight License (Mistral Large 4). Parameter counts for models that use a mixture-of-experts design show total and active parameters; see Mixture of Experts Explained.
What do open-weight licences allow?
Licence names group into three kinds. Read the licence file in the repository before use, because terms change between versions of the same family.
| Licence kind | Examples on this page | What to check |
|---|---|---|
| Standard permissive licence | Apache 2.0 (gpt-oss-120b, Mistral Small 4, Qwen3.8-27B) | Notice and attribution duties; Apache 2.0 includes a patent grant. OpenAI's page describes Apache 2.0 as allowing building "without copyleft or patent restrictions". |
| Model-specific open licence | OpenMDW-1.1 (Nemotron 3 Ultra) | Scope covers weights, data and code; compare the text against your legal team's approved list |
| Publisher's own licence | Kimi K3.1 Pro License, Mistral Ecosystem Open Weight License | Commercial use conditions, attribution, user or revenue thresholds, acceptable use clauses and redistribution rules, all of which are set by that publisher |
A useful habit is to record the licence name, version and repository commit for every model you deploy, because the same family can move between licences across releases.
How much does each option cost?
Closed models are billed per million tokens. As examples, on 2026-11-15 the providers' pricing pages listed Claude Fable 5.2 at $8 input and $40 output, GPT-6.1 Sol at $2 and $10, and Grok 4.7 at $2 and $6 (USD). Many open-weight models are also sold as hosted APIs: Mistral's card lists Mistral Large 4 at $0.70 input and $2.10 output, and DeepSeek lists deepseek-flash at $0.30 input (cache miss) and $1.20 output at peak hours. See Model Pricing Explained for how token billing, caching and batch discounts work.
Self-hosting replaces token fees with hardware. NVIDIA's card for the BF16 Nemotron 3 Ultra checkpoint lists a minimum of 8 GB200, B200, GB300 or B300 GPUs, 16 H100 or 8 H200, and links a smaller NVFP4 variant. OpenAI says gpt-oss-120b "fits into an H100 GPU". Smaller models run on a workstation; see Running LLMs Locally Explained.
Where does each model run?
- Closed models: the provider's API and apps, plus named cloud marketplaces. Anthropic and OpenAI document Amazon Bedrock model IDs for some models, for example global.anthropic.claude-fable-5-2 and openai.gpt-6.1-sol.
- Open-weight models: anywhere the licence permits. Gemma downloads are offered on Kaggle and Hugging Face. gpt-oss-120b is also listed on OpenAI's API endpoints and on Amazon Bedrock, so an open-weight model can be consumed exactly like a closed one.
- Air-gapped or sovereign environments: only open-weight models can run with no network path to the publisher.
How do teams choose between open-weight and closed models?
The criteria below are factual questions, not a verdict. The answer differs by workload, and many organisations use both through an AI gateway. See Self-Hosted LLMs vs Hosted LLM APIs and Foundation Model Selection Guide.
| Criterion | Question to answer | Points towards |
|---|---|---|
| Data location | Must prompts and outputs stay inside your network or jurisdiction? | Open weights self-hosted, or a closed model with a documented regional option (for example Anthropic's inference_geo or OpenAI's regional processing, both priced at a 10% uplift) |
| Licence fit | Does the licence allow your use, user count and redistribution? | Whichever licence your legal team approves |
| Version control | Do you need a model that never changes underneath you? | Open weights pinned by file hash, or a closed model's dated snapshot until its published retirement date |
| Operations capacity | Can you run GPU serving, scaling and patching? | Closed API if not |
| Customisation | Do you need full fine-tuning or weight-level changes? | Open weights |
| Volume and cost | Does steady high volume justify owned hardware? | Compare a token bill against hardware and staff cost for your actual volume |
| Access restrictions | Is the model you want restricted to approved customers? | Check the provider's access status; restricted closed models need an approval process |
What security and governance issues differ?
The controls move, they do not disappear.
- Model file supply chain (open weights): downloaded weights and tokenizer files are third-party artefacts. Verify the publisher's organisation on Hugging Face, pin a commit, scan files and record provenance. See the AI Model Scanning and Supply Chain guide and Model Signing vs Model Scanning.
- Credentials (closed models): API keys for hosted models are high-value secrets that attackers resell. See the LLM Provider API Key Security and LLMjacking Guide.
- Data retention: closed providers document their own retention and training-use terms; Anthropic's Claude Fable 5.2 page, for example, states that the model requires 30-day data retention and is not available under zero data retention unless Anthropic expressly authorises it. Self-hosted models keep whatever your logging keeps. See the AI and Privacy Guide.
- Safeguards: a closed provider's classifiers run server-side. Open weights carry whatever safety training the publisher applied, and fine-tuning can change it, so add your own input and output controls.
Common misunderstandings
- "Downloadable means free for any use": custom licences can restrict commercial use, require attribution or set thresholds. Read the licence file, not the model card summary.
- "Open-weight means I can see the training data": data release varies by publisher; most releases describe data at summary level only.
- "Closed means no control over versions": closed providers publish dated snapshots and deprecation schedules; plan migrations against those dates. See LLM Model Naming and Versioning Explained.
- "Open-weight is always cheaper": idle GPUs, staff time and energy are real costs; at low volume a hosted endpoint can cost less.
- "Parameters tell you the hardware you need": with mixture-of-experts models the total parameter count sets memory needs while active parameters drive compute per token; quantisation changes both.
Practitioner checklist
- Record for each model: provider, model ID or repository, open or closed, licence name and version, and date checked.
- Have legal review every custom licence before production use.
- Pin open-weight models to a repository commit and file hash; store them in an internal registry.
- Scan downloaded model files and verify the publishing organisation.
- For closed models, track each model ID's published deprecation and retirement dates.
- Confirm data retention and training-use terms in the provider's documentation for each closed model.
- Route all model traffic, open and closed, through one gateway for logging and access control. See the AI Gateway and LLM Gateway Guide.
- Re-check licences and access status quarterly; both change between releases.
Standards and references
- Open Source Initiative: The Open Source AI Definition 1.0 (accessed 2026-11-15)
- OpenAI: gpt-oss-120b model page (accessed 2026-11-15)
- Google: Gemma documentation (accessed 2026-11-15)
- Mistral AI: Mistral Large 4 model card (accessed 2026-11-15)
- Mistral AI: Mistral Small 4 model card (accessed 2026-11-15)
- Alibaba Qwen: Qwen3.8-27B model card (accessed 2026-11-15)
- Moonshot AI: Kimi K3.1 Pro model card (accessed 2026-11-15)
- NVIDIA: Nemotron 3 Ultra 550B-A55B model card (accessed 2026-11-15)
- DeepSeek: Hugging Face organisation page (accessed 2026-11-15)
- Meta: developer site (accessed 2026-11-15)
- Anthropic: Claude Fable 5.2, Claude Mythos 5.2 and Pricing (accessed 2026-11-15)
- OpenAI: GPT-6.1 Sol model page and API pricing (accessed 2026-11-15)
- xAI: API pricing (accessed 2026-11-15)
- DeepSeek: Models and pricing (accessed 2026-11-15)
Related NHI Mgmt Group resources: Build vs Buy vs Fine-Tune AI Models, Small Language Models vs Large Language Models, Model Cards and System Cards Explained, AI Model Security Guide.