Open weight models are models whose trained parameters are published for download, so you can run them on your own hardware or pick any host that serves them, under a license that says what you may do with them. That makes them portable in a way a closed API model is not. It does not make them open source, and it does not make every license permissive.
This guide covers what "open weights" gives you, the licenses attached to the main families as of 19 October 2026, a table of what each family publishes and where it can run, and when open weights are the right choice.
Open weights vs open source: what you actually get
An open weight release gives you the parameters: the files a serving engine loads to run the model. Usually it also gives you a model card, a chat template and inference examples. It rarely gives you the training data or the code that produced the weights.
The Open Source Initiative's Open Source AI Definition 1.0 sets a higher bar. It requires the "preferred form for modification" in three parts: detailed information about the training data, the complete source code used to train and run the system, and the model parameters. It also requires four freedoms: to use the system for any purpose without asking permission, to study it, to modify it, and to share it1.
Most of the models in this guide meet the parameters requirement and, under a permissive license, the freedoms. Few publish training data information in the depth the definition asks for. "Open weights" is the accurate term, and it is the one this guide uses.
What open weights change in practice:
- You choose the host. The lab that trained the model and the company that runs it can be different parties, so the data path is a hosting decision, not a model decision. This is the core of running Chinese open-weight models in production.
- You control versions. A checkpoint you downloaded does not change underneath you. A hosted API can.
- You can fine-tune. The weights are the starting point for your own variant, subject to the license.
- You take on what a closed API hides. Serving, capacity, upgrades and license compliance become yours, or your host's.
The licenses, from permissive to conditional
Read the license attached to the exact model you deploy. It can differ between families, and between models inside one family.
Apache 2.0 and MIT
The simplest case. gpt-oss-120b and gpt-oss-20b are Apache 2.023. So are Qwen3.8-27B4, Gemma 4 31B5, Mistral Small 46 and Mistral Large 37. DeepSeek V4.1 Flash uses the MIT License, which its card says covers both the repository and the model weights8. These allow commercial use, modification and redistribution, with notice requirements and no warranty.
Community licenses with thresholds
Meta's Llama models use custom community licenses. The Llama 3.3 Community License requires a separate license from Meta if your products exceed 700 million monthly active users, a visible "Built with Llama" notice, model names beginning with "Llama" for models trained on Llama outputs that you distribute, and compliance with an acceptable use policy9. Llama 4 Scout carries the Llama 4 Community License with the same 700 million user threshold10.
MIT-style licenses with commercial conditions
Several 2026 releases start from MIT's permission text and add conditions:
- Kimi K3 (Moonshot AI): a company offering third parties API access to the model's inference or fine-tuning needs a separate agreement once its revenue passes $20 million over any 12 months; products above 100 million monthly active users or $20 million in monthly revenue must display "Kimi K3" in their interface. Internal use that exposes nothing to third parties is exempt11.
- GLM-5.3 (Z.ai): a model-as-a-service business with more than $10 billion in revenue over 12 months must pass Z.ai's security review before commercial use; embedding the model in product features or relaying requests to models hosted by others does not count12.
Licenses with notices and prohibited uses
The MiniMax Community License for MiniMax M3 is the most restrictive of those read for this guide. Commercial use requires a visible "Built with MiniMax M3" notice, a one-time notice to MiniMax at or below $20 million in yearly revenue and prior written authorization above it. It also prohibits specific uses, including any military purpose13.
Qwen's largest model, Qwen3.8-2.4T-A95B, is listed under a license named "qwen3.8-max" rather than Apache 2.014. The smaller Qwen3.8 models and the largest one are not under the same terms, which is exactly why the license check belongs to the model, not the family.
The main open weight families as of October 2026
The table lists one or two current models per family whose model card was opened for this guide. Sizes are as each card states them; "active" applies to mixture-of-experts models, which load all parameters but use a fraction per token.
| Family | Model | License | Size | Context | Where it can run |
|---|---|---|---|---|---|
| gpt-oss (OpenAI) | gpt-oss-20b | Apache 2.0 | 21B total, 3.6B active | — | Within 16 GB of memory; vLLM, Ollama, Transformers3 |
| gpt-oss (OpenAI) | gpt-oss-120b | Apache 2.0 | 117B total, 5.1B active | — | One 80 GB GPU such as H100 or MI300X2 |
| Llama (Meta) | Llama 3.3 70B Instruct | Llama 3.3 Community | 70B dense | 128K | More than one 80 GB GPU at 16-bit9; Groq serves it15 |
| Llama (Meta) | Llama 4 Scout | Llama 4 Community | 109B total, 17B active | 10M | One H100 with int4 quantization10 |
| Gemma (Google) | Gemma 4 31B | Apache 2.0 | 31B dense | 256K | vLLM, SGLang; quantized builds for llama.cpp, Ollama5 |
| Qwen (Alibaba) | Qwen3.8-27B | Apache 2.0 | 27B dense, with vision | 262,144 native | SGLang, vLLM4 |
| Qwen (Alibaba) | Qwen3.8-2.4T-A95B | qwen3.8-max | 2.4T total, 95B active | 262,144 native | Multi-node cluster or a host14 |
| DeepSeek | DeepSeek V4.1 Flash | MIT | 763B listed, FP8 weights | Up to 1M | vLLM, SGLang8 |
| Mistral | Mistral Small 4 | Apache 2.0 | 119B total, 6.5B active | 256K | vLLM6 |
| Mistral | Mistral Large 3 | Apache 2.0 | 675B total | 256K | One 8×H200 or B200 node at FP87 |
| Kimi (Moonshot AI) | Kimi K3 | Kimi K3 License | 2.8T total, 104B active | 1,048,576 | vLLM, SGLang on a multi-GPU cluster16 |
| GLM (Z.ai) | GLM-5.3 | GLM-5.3 license | 753B listed | Not stated on card | SGLang, vLLM, KTransformers17 |
| MiniMax | MiniMax M3 | MiniMax Community | ~428B total, ~23B active | 1M | SGLang, vLLM18 |
Three notes on reading it:
- Llama's newest checkpoints are older than the rest. The most recently updated language models on Meta's Hugging Face organization are the Llama 4 releases, last updated in May 202519. Every other family in the table shipped in 2026.
- Gemma 4 spans the range. Besides the 31B dense model, the collection holds a 26B mixture-of-experts model with about 4B active parameters and E2B and E4B models aimed at edge devices20.
- Precision is part of the release. gpt-oss ships MXFP4 weights for its experts, which is why the 120B model fits one 80 GB GPU2; Kimi K3 was trained with MXFP4 weights and MXFP8 activations16; DeepSeek V4.1 Flash publishes FP8 weights8. A host that serves a different precision is serving a different numerical model.
Self-host or use a hosted provider?
The weights are the same either way. What changes is who runs them.
| Self-hosted | Hosted inference provider | |
|---|---|---|
| Data path | Your infrastructure | The host's infrastructure and retention policy |
| Cost structure | Hardware and operations, paid whether traffic arrives or not | Per token or per request, no idle cost |
| Version control | You decide when to upgrade | The host decides what it serves and when it changes |
| Limits | What your memory allows | What the host configures for your account |
| License compliance | Entirely yours | Shared with the host under its terms |
| Who pages at 3 a.m. | You | The host, for its part |
Hosting options are broad. Hugging Face's Inference Providers documentation lists partners including Baseten, Cerebras, DeepInfra, Fireworks, Groq, Novita, Nscale, OVHcloud, Scaleway, Together and Z.ai behind an OpenAI-compatible router, and describes automatic failover between them when one is flagged unavailable21. Most labs also run their own API for their models. For self-hosting, every engine above exposes an OpenAI-compatible server; self-hosted models behind an OpenAI-compatible endpoint walks through vLLM in production. The infrastructure side of running them in your own data centre is covered in on-prem LLM.
Same weights, different hosts
A hosted open weight model is the model plus the host's choices. Groq's model list shows it plainly: it serves qwen/qwen3.8-27b with a 131,072-token context window and marks it Preview, meaning evaluation only and subject to discontinuation at short notice15. The model card states 262,144 tokens natively4. Neither is wrong; they are different products built on the same weights.
What to check for every host of a model you plan to use:
- Context and output limits as the host configures them, not as the card states.
- Precision. BF16, FP8 and lower formats are not interchangeable for every task.
- Chat template and tool-call parsing. A wrong template or parser turns tool calls into plain text.
- Status. Preview, beta and production carry different promises.
- Version pinning. Whether the host's model id points at a fixed checkpoint or a moving one.
Then measure. Same open-weight model, different hosts publishes a protocol and harness for checking whether two hosts behave the same, starting from run-to-run noise on one host. Pair it with evaluating on your real workloads so the comparison uses your prompts, not a benchmark's.
When open weight models are the right choice
Open weights earn their place when one of these is true:
- The data path must be yours. Regulated data, customer contracts or internal policy rule out sending prompts to a third party. Self-hosting is the only option that keeps every byte on your infrastructure.
- You need a fixed version. A workflow validated against one checkpoint should not change because a provider updated its model.
- You want a way out. A model several hosts serve gives you an alternative when one host has an outage, changes terms or retires the model.
- You will fine-tune. Domain adaptation, distillation and specialised variants start from weights.
- Steady, high volume makes owned capacity pay. Utilisation decides it; idle GPUs are expensive.
They are a weaker fit when you need the strongest available model on a hard task and a closed model wins your evaluation, when your traffic is spiky and small, or when nobody on the team will own a serving stack or a license review.
Running open weight models through INFRO
INFRO is a control plane over provider accounts you already hold. For open weight models that means connecting the hosts you chose with your own keys, such as Groq, Together AI, Fireworks AI or a lab's own API, and any server you run that speaks the OpenAI API as an OpenAI-compatible endpoint (connect a provider). Each host bills you directly; INFRO charges a subscription and takes no share of model spend.
A route can then put the same model on two hosts, or a self-hosted server and a hosted one, with failover between them (routes). Model allowlists keep production to the models and licenses your review approved, and every request record names the provider and connection that served it, with its latency and estimated provider cost, so a host-by-host comparison comes from your own traffic. INFRO does not make two hosts behave the same: whether a fallback host is equivalent is still a question your evaluation answers.