Open Weight Models: Licenses, Hosting and Business Control

Open weight models give a business the option to run and adapt a model under its own control, with the permissions set by its license. Meta's October 15, 2026 release of Muse Spark Open adds another US example: downloadable weights create a deployment choice, not an automatic saving or unrestricted commercial rights.

Start with an API unless a specific control requirement or a measured workload justifies operating the model yourself. An API is an interface your software calls to request a service; that service can run an open-weight model. You can evaluate the model through a host before deciding whether owning its deployment is worth the work.

What Are Open Weight Models?

Open-weight models let you obtain the trained numerical parameters that make a model produce answers. Those parameters are the weights. With compatible software and enough hardware capacity, you can load them and run inference, which means generating a response from an input. The original training data and training code may remain unavailable. Stanford HAI's definition makes that distinction explicit.

Think of receiving a mature plant from a nursery. You can move it into your greenhouse and tend it yourself. You have not necessarily received the nursery's complete cultivation records or the method that produced it. The downloadable model is the trained result; reproducing its original training is a different undertaking.

For an operator, the immediate gain is choice over where a particular model runs. Depending on its license, you may also modify it, fine-tune it on examples of your work, or distribute a derivative. Fine-tuning means further training the model to change its behavior; it is different from attaching documents to a prompt.

Consider an internal supplier-document assistant. Open weights could let your company keep the inference service inside its own infrastructure and retain a particular model version. They do not supply the document permissions, search system, user interface or error-handling workflow. You still have to build or buy those components.

Open Weight vs Closed Weight Models

The difference is whether you can obtain and operate the weights yourself. With a closed-weight service, you use the provider's interface under its terms; you cannot take that checkpoint to another server. With an open-weight release, a compatible deployment can run elsewhere, subject to its license.

This separates two questions that often get mixed together: who has the weights, and who operates your endpoint? A hosted API is an access method. It does not tell you whether the underlying model is open or closed.

Moving an application between hosts also takes work. Prompt formats, supported tools, media inputs, output parsing and serving settings can differ. Weight availability provides an alternative to the original provider; it does not make every endpoint interchangeable.

Open Weight Models vs Open Source

A download link establishes access to weights; it does not establish that the whole AI system is open source. A familiar software license on a checkpoint is useful evidence about permissions, but training transparency is a separate question.

The Open Source Initiative's Open Source AI Definition requires the freedom to use, study, modify and share, together with the materials needed to make modifications: parameters, the code used to train and run the system, and sufficiently detailed information about training data.

There is a nuance here: the definition does not simply demand that every original training record be publicly downloadable. It specifies information about data, including unshareable data, and where obtainable datasets can be found. Saying “open source means every dataset is public” oversimplifies the standard.

For procurement, keep these checks separate:

• Availability: Can you obtain the exact checkpoint, the saved model you intend to deploy?
• Permission: Does its license permit your internal use, customer-facing product, redistribution or model-hosting business?
• Transparency: What code, data information and training artifacts are available for inspection or reproduction?
• Operations: Can your chosen infrastructure deliver acceptable results and response times?

A business deploying an invoice classifier may mainly need commercial permission and a private runtime. A research group trying to reproduce training needs much more. Neither should infer the other requirement from the word “open.”

Open Weight AI Models: What the Licenses Allow

The permission belongs to a specific release, not to Meta, OpenAI, DeepSeek, Qwen, Mistral or NVIDIA as a whole. The examples below explain the differences a business needs to recognize. They are not a ranking.

Verified October 16, 2026: the maker cards, release pages and license texts linked here. The matrix is a comparison of those documents, not a hands-on model evaluation. For Mistral Small 4, the maker card identifies Apache 2.0; the terms are read from Apache's original license text.

Meta Muse Spark Open: Commercial Use Depends on the Activity

Meta released Muse Spark Open on October 15, 2026, adding a new US-developed model with downloadable weights. Meta describes a multimodal model for coding and agentic work, meaning workflows in which a model uses tools across multiple steps. Its release announcement offers both downloadable weights and named API hosts.

The Muse Model License permits use, modification, fine-tuning, distillation, derivatives, distribution and sales subject to its conditions. Keep the copyright notice, license and Usage Policy with covered copies, follow that policy, and prominently display “Built with Muse Spark Open” in the interface or accompanying documentation, website or settings.

Its derivative definition includes AI models improved using the software's outputs. That matters if your plan is to train another model from generated examples.

The separate-agreement condition targets Model as a Service businesses: when aggregate gross revenue across the licensee and affiliates exceeds $20 million during any consecutive 12 months, they must notify Meta and enter a separate written agreement before offering the model or derivatives through that business. Section 3 excludes internal use, research, other commercial activity, and features of a product that is not Model as a Service.

A private support assistant and a business selling model access therefore require different classifications. Neither escapes the remaining license conditions. Our Muse Spark Open deployment guide covers the hardware and hosting details.

OpenAI gpt-oss: Permissive Weights Still Need the Right Runtime

OpenAI gpt-oss-20b and gpt-oss-120b are downloadable reasoning models with Apache 2.0 licensing. OpenAI's maker card describes fine-tuning and configurable reasoning effort, which changes how much reasoning the model performs before answering.

The license file permits commercial reuse, modification and redistribution. Distribution requires a copy of the license, notices identifying modified files, and retention of applicable attribution and other notices, including NOTICE material where supplied. Apache also provides a defined contributor patent grant; that is not a blanket guarantee against every patent claim.

The operational limitation is concrete: OpenAI says these models must use its Harmony response format. A compatible runtime or chat template handles that structure. Downloading weights into an arbitrary chat wrapper does not establish correct behavior.

For a builder, this is a manageable integration dependency. For a buyer, it means the deployment proposal should name the runtime and checkpoint, not just promise “OpenAI running privately.”

DeepSeek: MIT Permission Does Not Supply a Managed Service

DeepSeek-V4-Pro provides a downloadable example with MIT terms. The maker card describes the V4 series as a preview, links the Pro and Flash checkpoints, and supplies local deployment guidance. These are specific artifacts examined here, not a claim that every DeepSeek release is identical.

The card explicitly applies the MIT License to the repository and model weights. The text permits use, copying, modification, distribution, sublicensing and sales, while requiring the copyright and permission notice in copies or substantial portions. It supplies no warranty.

For an internal document-processing service, those permissions remove a licensing obstacle to operating and modifying that checkpoint. They do not provide a response-time commitment, an incident team or an assurance that a differently named hosted endpoint serves the same version. Put those requirements in the service or deployment specification.

Qwen: The Family Name Hides a License Split

Alibaba's Qwen3.8-27B and Qwen3.8-2.4T-A95B have different licenses and input capabilities. The Qwen3.8-27B card describes a multimodal release, and its license file is Apache 2.0.

The larger checkpoint's Qwen3.8-Max License permits hosting, deployment, fine-tuning, derivatives and distribution subject to conditions. Retain the copyright and permission notice. A commercial product using it must prominently display the model name in its interface when that product exceeds 100 million monthly active users or $20 million in monthly revenue.

A separate condition applies when the licensee or affiliates conduct a Model as a Service or AI Work Assistant business and their aggregate revenue exceeds $50 million in any consecutive 12 months: obtain a separate license before commercial use. The internal-use exception requires that the model, outputs and underlying capabilities are not made available to third parties.

The definitions matter. AI Work Assistant focuses on independent products mainly for coding or office productivity, with exclusions for single-purpose tools, other-domain assistants and features of products with another primary purpose. Classify the actual product against the full text.

There is also a capability mismatch to catch: the large checkpoint's card says it is text-only and requires thinking mode, while the hosted Qwen3.8-Max version has additional features, including vision. A hosted trial cannot prove feature parity with the download.

Mistral: Use an Available Checkpoint as the Evidence

Mistral Small 4 is an available multimodal checkpoint whose maker card specifies Apache 2.0. The Mistral-Small-4-119B-2603 card describes text and image input, text output, instruction and reasoning modes, and customization through fine-tuning.

The Apache 2.0 text permits commercial modification and redistribution with its license, change-notice and attribution requirements. It does not require publishing every modification, and it does not grant unrestricted use of the maker's trademarks.

Do not carry that conclusion forward by brand association. Mistral's Large 4 announcement describes a public preview and says weights will be released by the end of the month. On this verification date, that announcement is a future commitment, not evidence of an available checkpoint or its download license.

An operator planning an immediate private deployment needs a release that can be obtained now. The Mistral Large 4 page covers the preview separately.

NVIDIA Nemotron: Read the Current Card's Governing Terms

NVIDIA Nemotron 3 Ultra uses OpenMDW 1.1 in the examined checkpoint. NVIDIA's model card describes a reasoning, conversational and agentic model ready for commercial and noncommercial use. Older Nemotron licensing pages are not a substitute for that card's governing-terms link.

The OpenMDW 1.1 license broadly permits dealing in the supplied model materials. Distributing them requires the agreement and applicable copyright and origin notices. It imposes no restrictions or obligations on using, modifying or sharing generated outputs.

The license also includes termination provisions for specified patent or copyright litigation, disclaims warranties, and leaves clearing other applicable rights to the user. Broad permission is not an indemnity promise.

NVIDIA describes released training data and recipes, while the same card also identifies private datasets. This illustrates why “open-weight” is a minimum description of access, not a claim that every maker releases exactly the same supporting material. If reproducibility matters, inspect the named artifacts rather than treating the label as a transparency score.

Where You Run It Changes the Business Trade-Off

Choose the deployment boundary before buying hardware. The same downloadable model can support quite different operating arrangements. The useful question is which responsibilities you want to own.

Hosted API: Keep the Service Work With a Provider

A hosted API is the fastest starting point when you need to establish task quality. You send a request to a service that loads and operates the model. Meta names Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as Muse Spark Open launch platforms; that list does not establish identical features or contractual terms across them.

For a funded founder prototyping a support assistant, hosting avoids a capacity commitment before the workflow is proven. Compare the exact input types, model revision, tool support, retention terms and region of the endpoint you will use.

The limit is data control. Open weights at somebody else's endpoint still mean sending prompts to that endpoint. Your agreement and the service configuration determine what happens there. An open model name does not turn a hosted call into local processing.

Your GPUs: Own the Deployment and Its Responsibilities

Self-hosting lets you decide how the inference service is operated. You can use machines you own or rent GPU capacity in a cloud account. GPUs are processors used to run the model computations. A runtime is the software that loads the checkpoint and serves requests; the model makers' deployment instructions should determine which runtime and format you select.

For a mid-market CTO with a requirement to keep supplier documents inside a controlled environment, this can be the deciding benefit. The boundary still needs to include document retrieval, request logs, monitoring, backups and any tools the model calls. A private model that sends document contents to an external search service has not met that requirement.

The responsibility also becomes yours: access control, patching, capacity planning, rollout testing, incident response and recovery. Assign an owner before introducing the dependency. A GPU reservation alone is not a production operating plan.

Open Weight LLM Models on a Local Device

Local deployment is useful when one person's machine can handle the workload and isolation matters more than shared scale. LLM means large language model. A solo technical builder might use a local model to classify files or assist with a repository while retaining control of the runtime.

OpenAI's gpt-oss card documents local paths, while Mistral Small 4's card lists local and server runtimes. Compatibility must be checked for the precise checkpoint and format. A model appearing in a catalog does not establish that it will fit your device.

A personal workstation also has ordinary availability limits: it can sleep, restart or be occupied by other work. Keep that distinction clear before turning a successful personal experiment into a shared service. Our best local LLMs guide handles the hardware-oriented selection question.

Open Weight Large Language Models: The Limits That Matter

Memory fit is the first hardware gate; usable throughput is the next. OpenAI says gpt-oss-20b can run within 16GB of memory using its shipped MXFP4 quantization, a lower-precision representation of weights. That is a maker-stated configuration, not a promise about every computer or request load.

At the other end, Meta's Muse Spark Open card states 700GB of GPU memory for NVFP4 and 2TB for BF16, two weight formats with different numerical precision. Its separate desktop claim refers to two Apple M5 Ultra Mac Studios using MLX, according to the announcement. “Runs locally” spans very different classes of equipment.

The weight files are not the entire running workload. Input history, intermediate computation and concurrent requests consume resources too. Measure the response time and successful throughput at your actual prompt lengths. Neither a parameter count nor a maximum context window tells you how many customers your service can support acceptably.

What Changes for Builders, Operators and Buyers

The business case is strongest when control solves a named problem and somebody can operate the service. Volume can strengthen that case, but it does not settle it by itself.

For builders, the benefit is an inspectable deployment choice. You can keep a selected artifact and change serving settings or fine-tune when the license allows it. The work is making the application behave correctly around that artifact: formatting requests, interpreting tool calls and handling failure. Start with a bounded workflow such as extracting supplier identifiers, with explicit examples of acceptable and unacceptable output.

For operators, the benefit is control over the processing boundary and change schedule. The corresponding work is documenting where data goes and keeping the system healthy. A senior support operator should require a clear handoff for unreadable attachments, uncertain classifications and unavailable inference. Model ownership does not eliminate the need for human escalation.

For buyers, the benefit is a credible alternative deployment path. Require the proposal to name the model revision, license, infrastructure, support responsibilities and migration procedure. “Powered by open models” is insufficient if you cannot obtain the configuration or move the application without rebuilding its surrounding components.

When Volume Helps, and When It Does Not

Self-hosting becomes economically plausible when useful work keeps the capacity busy. A predictable document-processing queue can often use capacity differently from a customer-facing assistant that needs immediate responses during irregular bursts.

Compare the whole service: compute, storage, monitoring, engineering time, idle capacity, retries, fallbacks and any applicable licensing agreement. Then compare accepted results at the latency your workflow requires. A cheaper generated response is not a saving if it creates more review work or repeats a failed task.

A founder with uncertain demand should learn on a hosted service before making a hardware commitment. A CTO with sustained load and existing infrastructure expertise has a stronger reason to measure a private deployment. Neither needs a universal token-volume threshold invented from someone else's configuration.

This is the operating distinction: an API provider absorbs much of the capacity and service-management work; a self-hosted deployment lets you control more of it while taking responsibility for it.

What Is Overhyped About Open Weights?

The overstatement is that access to weights automatically delivers freedom, privacy and savings together. Each has its own condition.

“Open means unrestricted.” The Muse and Qwen custom licenses show why this fails. A download can permit many uses while attaching conditions to branding, derivatives or particular businesses. A permissive license also has terms worth preserving when you distribute a model.

“Open means private.” Privacy follows the deployed system's data flows. Check the inference host, retrieval service, tools, logs and administrators. Running the model on your own server addresses one part of that map.

“Open means the API product, downloadable.” Qwen's card explicitly distinguishes its large text-only checkpoint from the hosted version's additional features. Applications also add orchestration, tools and interfaces around a model. Taking weights does not reproduce all of that behavior.

“Open means no recurring work.” Hardware still needs capacity, software still changes, and incidents still need owners. A team can gain autonomy and still spend more overall. Choose autonomy when it solves a business constraint; claim savings only after measuring them.

“Announced open weights are available weights.” Mistral Large 4's release timetable is the practical counterexample. Plan an evaluation around an actual artifact and its published terms, not an expected download.

What to Do Differently on Monday

Evaluate one workflow before changing your infrastructure. Act now if a data boundary, customization requirement or dependency on a single service is already causing trouble. Wait on hardware if the model's task quality or workload is still uncertain. An operator satisfied with an existing API and its terms has no automatic reason to migrate.

1. 1

   Name the requirement

   Choose a concrete job and state why the current arrangement is insufficient: documents must stay inside a specified environment, a model revision must remain available, or a stable processing queue needs evaluation. Define acceptable output and escalation behavior.

2. 2

   Identify the artifact and its permissions

   Save the checkpoint identifier, revision, model card and license. Describe whether you will use it internally, expose a product feature, redistribute it or sell model access. Resolve the applicable conditions before building around them.

3. 3

   Evaluate through the permitted route

   Use a host if your data requirement allows it; otherwise evaluate in the required environment. Compare representative tasks with the current service. Record errors, response time, review effort and any endpoint features the application relies on.

4. 4

   Assign operations before moving production

   Name who owns capacity, access, monitoring, updates and incidents. Require a rollback path and a repeatable deployment. Buy hardware only after the model, workload and operating plan justify it.

Use the best open-source LLMs guide for model selection after these requirements are clear. Here, the decision is what control you need and what responsibility you are prepared to accept.

Frequently Asked Questions

What is an open weights model?

It is a trained model whose weights are available to obtain and run with suitable software and hardware. Its training data and training code may remain unavailable. The license determines which uses, modifications and distributions are permitted.

What's the best open weight model?

There is no task-independent answer. First establish the permitted use, required inputs, hardware or host, and acceptable output. Use the linked model-selection guide to shortlist candidates, then evaluate your workflow. This page compares access and responsibilities rather than ranking performance.

What is the difference between open and closed weight models?

Open weights can be obtained for a compatible deployment under their license. Closed weights remain with the provider. Either kind can be accessed through an API; using an API does not tell you whether the underlying model is downloadable.

Are open weight models free?

A license may grant use without a royalty, but operating a model still consumes hardware and staff time. A hosted provider can charge for inference, and a custom license can require a separate agreement for specified businesses. Check both the model's terms and the service you will use.

Will open weight models be banned?

A release announcement cannot establish future legal permission. Assess the specific model, intended use and deployment jurisdictions against the applicable rules when you deploy. Do not make a migration decision on a prediction that all open-weight models will be either prohibited or permanently unrestricted.

How do open weight models make money?

Downloadable weights can sit alongside paid hosting, support, customization and commercial licensing arrangements. The weight download and the service around it are different offerings. Muse and Qwen also illustrate conditional separate-license requirements for certain businesses.

How far behind are open weight models?

The label does not specify a performance gap. Compare exact releases on your task, including the serving configuration and features available at the endpoint. An aggregate benchmark cannot establish your error rate, throughput or review burden.

Are open weight models safe?

Weight availability does not establish safe application behavior. Evaluate failures in the intended workflow, restrict tool permissions, protect sensitive inputs and define human escalation. Meta's model card itself calls for application-specific safety testing and additional safeguards for agents taking consequential actions.

For practical AI adoption decisions, join the newsletter.