What Does "Open Weight" Mean?
An open-weight model is one whose trained parameters — the weights — are published so anyone can download them. Once you have the files, you can load the model in compatible inference software such as vLLM, SGLang, Transformers, or llama.cpp, run it on your own hardware or a cloud instance, and often fine-tune it for a specific task.
Open weight does not mean restriction-free. The publisher still attaches a license that controls what you can do with the weights, how you can distribute derivatives, and whether commercial use requires extra steps. Some licenses are familiar, like Meta's Llama 3 license, while others are custom, like the Kimi K3 License.
The key point is accessibility: the weights are out in the open, so users and researchers can inspect, self-host, and experiment without asking the publisher for permission each time. That openness is what the term open weight specifically describes.
What Does "Closed Weight" Mean?
A closed-weight model keeps its weights inside the organization that built it. You do not download the parameters; you interact with the model through an API, a web interface, or another hosted service that the provider operates.
This model shifts most of the operational burden to the publisher. You do not choose the inference engine, patch the runtime, provision GPUs, or estimate memory usage. You pay for usage — usually per token, per request, or through a subscription — and the provider handles scaling, updates, and availability.
The trade-off is less control. You cannot see the exact weights, fine-tune them directly, or run the model in an air-gapped environment unless the provider offers a separate private-deployment tier. For many teams, that trade-off is acceptable because the managed experience is faster to adopt and requires no specialized machine-learning infrastructure.
Open Weight vs Closed-Weight: The Key Differences
The practical gap between open-weight and closed-weight models comes down to who controls the artifact and the runtime. The table below summarizes the six dimensions that usually drive the decision.
| Dimension | Open-weight model | Closed-weight model |
|---|---|---|
| Weight access | Weights are publicly downloadable | Weights are kept private by the provider |
| Deployment | Self-hosted or run on any compatible inference engine | Available only through the provider's API or hosted service |
| Fine-tuning | Can be fine-tuned locally if the license permits | Fine-tuning is available only through provider tools, if offered |
| Privacy and data control | Can stay inside your infrastructure when self-hosted | Usually sends input to the provider's servers |
| Cost structure | Up-front hardware and operations cost; no per-token meter | Usage-based pricing; no hardware purchase required |
| Transparency | Weights can be inspected, but training data and code may not be public | Weights, data, and code are generally not public |
Examples of Open-Weight Models
Kimi K3 is the clearest current example relevant to OpenK3 readers. Moonshot AI released the K3 weights on Hugging Face at moonshotai/Kimi-K3. The repository contains 96 Safetensors shards, totals approximately 1.56 TB, and lists 2.8 trillion total parameters with 104 billion activated parameters. The context length is listed at 1,048,576 tokens. K3 is distributed under the custom Kimi K3 License, which permits personal and research use freely and allows commercial use subject to revenue and user thresholds.
Other well-known open-weight models include:
- DeepSeek-V3 and DeepSeek-R1 — released by DeepSeek with downloadable weights and permissive research use, subject to their license terms.
- Meta Llama 3 and Llama 3.1 — widely used open-weight families published by Meta with a custom license that requires acceptance and imposes some usage limits.
- Qwen2 and Qwen2.5 — Alibaba Cloud's open-weight model series, released under licenses that vary by model size and intended use.
Examples of Closed-Weight Models
Closed-weight models are common in consumer and enterprise AI products. Examples include OpenAI's hosted GPT models and Anthropic's hosted Claude models. In both cases, the weights are not distributed to end users; developers and consumers access capabilities through the providers' APIs, web applications, or partner integrations.
These models are typically updated on the provider's schedule, and users benefit from ongoing safety work, system prompts, and infrastructure scaling without managing the deployment themselves. The limitation is that you cannot download the weights, inspect their exact values, or run the model offline without a specific enterprise or private-deployment agreement.
Open Weight vs Open Source: Not the Same Thing
This distinction matters because the two phrases are often used interchangeably in casual conversation, but they mean different things to lawyers, researchers, and standards bodies.
Open weight only says that the trained weights are public. It does not promise that the training dataset, the training code, the exact hyperparameters, or the evaluation pipeline are also public. You can run the model, but you may not be able to reproduce it from scratch.
The Open Source Initiative's Open Source AI Definition uses a broader standard: the relevant data information, code, and parameters must be available so people can study, use, modify, and share the system under the stated terms. A model can be open weight without meeting that standard. Kimi K3 is therefore best described here as an open-weight model, not an open-source model, because the audited release publishes the weights and a custom license rather than a complete training-data and training-code package.
That distinction has real legal and practical consequences. Calling an open-weight model open source can lead teams to assume they have broader rights or reproducibility than the license actually grants.
How to Choose
The right choice depends on what you control and what you want to optimize for. Use the checklist below to narrow the decision quickly.
- Choose open weight if you need to self-host, keep data inside your network, fine-tune on private data, or build a derivative model. This is common in regulated industries, research labs, and cost-sensitive long-running services.
- Choose closed weight if you want an API that works immediately, need elastic scaling without hardware planning, or prefer that the provider handles updates, safety filters, and infrastructure maintenance.
- Read the license before committing. Open-weight releases often have custom terms, revenue thresholds, or display requirements that affect commercial use. Closed-weight services have their own terms of service, data-use policies, and rate limits.
- Remember that open weight and closed weight are not permanent categories. A provider may start closed and later release weights, or release weights while keeping the best version behind an API.