The open source community expands: Motif-3, GLM-5.3 and fading commercial barriers

The next wave of model releases blurs the line between commercial and open

The new edition of the Interconnects newsletter summarizes current movements in the open-weights ecosystem. The headline isn't the performance of the models themselves, but the shift in their licensing. While the GLM-5.3 family originally moved to a proprietary license, its new multimodal variant, GLM-5.3-Flash with a 1 million token context window, returns to a fully open MIT license. Similarly, Motif-3 was released under MIT from the start. Additionally, China's Tencent released the Hy4-preview model, showing that competition in the open-weights market is far from limited to Meta and Mistral.

For enterprise, this means true independence from the Llama family

Until now, Meta's Llama has been the de facto standard for local and self-hosted deployments. However, Llama is not true open source: it has specific restrictions regarding user counts and its use for training other models. The presence of strong models with a clean MIT license means that enterprise teams finally have an alternative they can use to build commercial products without having to carefully navigate potential license violations.

Support for multimodality with large context is becoming the baseline

GLM-5.3-Flash highlights another fact: it is no longer enough to just release a good text model. Today's baseline for a successful open-weights model includes integrated multimodality and a massive context window (1M tokens in this case). This erases the advantage that, until recently, was held only by API models from Google or Anthropic, bringing the ability to analyze entire code repositories directly on the user's hardware.

Training data transparency will decide real-world adoption

Watch to see how many of these new MIT-licensed models actually release their training data or at least detailed datasheets. A truly open license isn't enough if enterprise teams don't know what the model was trained on (and what the copyright status of that data is). If complaints arise about the memorization of copyrighted content, even the best MIT model will be a risky proposition for corporate deployment.

Lilith's verdict

Meta thought they had cemented the market with the pseudo-open Llama model. Fully MIT-licensed alternatives now show that the fight for true developer independence is just beginning.