Open-weight models: the other half of the language-model market

The open-weight story is no longer a side market for hobbyists. In August, Meta released Muse Glimmer, IBM shipped Granite 4.2, and DeepSeek and Qwen remained serious competitors outside the closed-API race. The question is shifting from whether downloadable models can compete to where they run, what their licenses permit, and whether governments can regulate files that anyone can copy.

Meta reopened the US race

Meta released Muse Glimmer's 30-billion-parameter weights on August 10 under Apache 2.0. It is designed for local agent workflows and can fit on a single consumer GPU, with integrations arriving for Ollama, LM Studio, llama.cpp, MLX, and other familiar tools. That is the practical pitch: download the model, run it near private data, and avoid a metered endpoint for every small task.

The bigger announcement was still pending. Meta said the weights for its frontier Muse Spark 1.2 would follow "soon," but had not published a date or license. Glimmer is real and downloadable now; Spark is a promise. Treating those as the same release is how open-weight buzz gets ahead of the evidence.

IBM made the enterprise case

On August 25, IBM released Granite 4.2 in 3B, 8B, and 30B sizes under Apache 2.0. The family adds switchable reasoning modes, long context, tool calling, and agentic reinforcement learning in the larger models. IBM also released small Granite Speech models aimed at local and edge deployment. The release is less about beating every frontier benchmark than offering a permissive license, a documented training story, and models an enterprise can actually host.

China keeps the pressure on

DeepSeek, Qwen, Moonshot, and other Chinese labs made open weights part of the geopolitical competition before Meta returned to the strategy. Their models are attractive because they can be downloaded, adapted, and served outside a US vendor's API or policy layer. OurChinese AI report covers the larger stack: domestic models, chips, cloud, and distribution.

Open weight is not open-source

Downloadable parameters do not automatically include training data, source code, evaluation results, or a permissive license. Apache 2.0 is straightforward. Other model licenses can limit commercial scale, redistribution, or certain uses. The name matters less than the actual model card and terms. A company that self-hosts also owns the security patching, abuse monitoring, evaluation, and upgrade schedule that a hosted provider handled before.

The policy fight is getting sharper

Washington is reportedly exempting open-weight releases from a new voluntary pre-release review aimed at closed frontier models. Supporters say that keeps US developers from surrendering the open ecosystem to China. Critics say it creates a safety loophole: once weights are public, a provider cannot recall them or monitor every downstream deployment. The same political divide now runs through procurement, export controls, and debates over whether open models should be treated as software, products, or infrastructure.

The point

Open-weight models help companies avoid API fees, vendor lock-in, sending data to another company, and depending on a foreign supplier. They aren't free: the cost moves to GPUs, hosting, security, testing, and legal review. Before choosing one, ask three questions: can you download the weights, does the license allow your use, and can your team run it safely? Meta's Glimmer and IBM's Granite make open weights more practical, but the license and running costs decide whether they are the better choice.