A model is mostly a very large file of numbers, its weights. When a developer publishes that file, anyone can run the model on their own hardware.
How it works
Training produces the weights: the parameters a neural network learned. The Open Source Initiative describes open weights as “the final weights and biases of a trained neural network.” With them, you can run inference, fine-tune or quantize the model.
What open weights usually do not include, per OSI:
- the training code;
- detailed information about the training data.
OSI’s Open Source AI definition requires both, alongside the weights, so others can study and reproduce the model. Many popular models are open-weight but not open source in that sense. US government reports use a related phrase: models with “widely available model weights.” Google, for example, calls its Gemma family “open models.”
Every open-weight release comes with a licence. Common ones include Apache-2.0 and MIT, while others restrict commercial use. The licence decides whether you may serve the model to other people through an API.
Why it matters
- Choice of where it runs. Self-host for full control, or call one of several providers that serve the same model.
- Privacy. You can keep data on your own hardware, or pick a host with zero data retention.
- No lock-in. If one provider changes terms, the same weights run elsewhere. An OpenAI-compatible API makes the switch a configuration change.
- Same name, different serving. Two providers can serve one model with different quantization, context limits or settings. Test on your own prompts.
On BroadNet Inference
BroadNet’s model catalogue says “~270 models run on our GPUs” (lab figure, September 2026). That counts every model BroadNet says it has run, benchmarked, quantized or fine-tuned, including ones it ruled out, not only the models on its API. Its privacy statement describes the hosting as BroadNet-owned or leased US hardware with partner capacity as backup. Its model catalogue shows licences, because, in its words, they “decide what can be served through an API.” BroadNet Gemma 4 VL 26B is BroadNet’s build of Google’s Gemma 4. BroadNet Transcribe Arabic is based on Cohere Transcribe, and BroadNet reports it at #2 of 37 (its September 2026 figure; the live board listed it #4 of 39 on 15 November 2026) on the Open Universal Arabic ASR Leaderboard, and first among models whose licences permit API serving.
To pick a model for Arabic text, see choosing an LLM for Arabic.