Meta’s Bold Open-Weight Bet is Meant to Muse Developers
By Siddharth Jindal ·
Meta is back to chasing Llamas. The company, once synonymous with open-weight models, has returned to its roots.
The company has released Muse Glimmer, a 30-billion-parameter open-weight model built for agentic tasks that can run locally on a consumer device with a single GPU. It will soon release the weights for Muse Spark 1.2, its most capable model yet.
Muse Glimmer is being released under the Apache 2.0 licence, with its weights available through Hugging Face. Meta says the model is intended for local agents, function calling, coding, and LLM-as-a-judge workloads without requiring a cloud connection.
Meta hopes to replicate the success it achieved with its Llama series. Developers could fine-tune the models, quantise them, and deploy on their own infrastructure, creating an ecosystem around Llama that extended far beyond Meta’s own products.
But Llama 4, launched a year ago, failed to generate the same momentum. Meta subsequently overhauled its AI efforts, bringing Alexandr Wang to lead Meta Superintelligence Labs, which launched its first model, Muse Spark, in April this year.
The timing with Muse Glimmer couldn’t have been better, as the developer community is seeking frontier open-weights to bring down token costs.
“It may not be called Llama anymore, but seeing Meta release a frontier open-weight model under an Apache license feels like history repeating itself,” Adithya S Kolavi, an AI researcher at Hugging Face, wrote on X. “Huge win for the open ecosystem.”
Kyle Chan, an analyst who covers China’s technology sector, believes Meta’s return to open-weight AI could put pressure on Chinese models. “An open-weight version of Muse Spark 1.2 would offer similar capabilities to top Chinese models, like Kimi K3 and GLM-5.2,” Chan wrote on X.
Moreover, he added that Meta’s US origins could give it an advantage, particularly among American and other Western users, who may prefer US models because of lower compliance and reputational risks and fewer unknowns.
“The new Meta open weight models are markedly different from Llama, from a licensing perspective, as they are shared on a true open source licence without any commercial restrictions,” Amanda Brock, CEO of tech non-profit OpenUK, tells AIM.
She further adds that the model is not fully open source as only weights are available under a licence, rather than providing access to the underlying training data, code and other components.
Challenging Chinese Models
In the last two months, US open-weight models have been under immense pressure from Chinese models.
Moonshot AI’s Kimi K3, a 2.8-trillion-parameter model with a 1-million-token context window, has delivered strong results across coding, reasoning, agentic, and vision tasks. At the same time, Z.ai’s GLM-5.2 has emerged as one of the strongest open-weight models, scoring 62.1% on SWE-bench Pro. DeepSeek’s V4 Pro and Alibaba’s Qwen3.6 are also among the strongest Chinese open-weight models.
Chinese AI companies have made open weights a major distribution channel, giving developers access to powerful models that can be downloaded, adapted, and deployed on their own infrastructure rather than relying on closed APIs.
China’s growing influence is already visible on OpenRouter, a widely watched model marketplace. Chinese models accounted for 63% of token usage by US companies in the first week of July, up from less than 10% a year earlier.
“None of the US models really compete with the Chinese open-weight models at this moment in time, but things change fast in this space,” says Brock
Zuckerberg has advocated for policies that make it easier for American developers to build and distribute open-weight models. At the same time, he believes banning Chinese models is not the way ahead.
“Restricting people and companies from using the leading open source models—wherever they come from—will reduce the quality of AI accessible to them, and centralise AI rather than putting more power in people’s hands,” he added in his blog post.
The US government has also come to the realisation that restricting open-weight models may hamper AI innovation. The Donald Trump administration has told AI companies that it would not subject open-weight models to the voluntary pre-release safety testing framework being developed by the White House for proprietary models that achieve state-of-the-art performance in cybersecurity and hacking benchmarks, according to The Wall Street Journal.
At the same time, Zuckerberg has also called for closer coordination between frontier AI companies and the government. “I think it will be more productive to have a close proactive collaboration between frontier labs and the government rather than a rigid process and review timeline that is followed in all cases,” he wrote in the blog post.
Meanwhile, Arpit Bhayani, creator of DiceDB, tells AIM that developers picking Muse depends more on geopolitics. “People or countries who do not mind using Chinese models will continue to use Kimi and GLM. But the moment geopolitics kicks in, that might be the biggest reason for people to switch.”
Meta Needs Developers
The return to open weights therefore looks like an attempt to rebuild Meta’s position in the developer ecosystem.
Meta does not necessarily need every developer to use Muse instead of GPT-5.6 or Claude. It needs them to keep Meta models in the conversation when they decide what models to build on, fine-tune, and deploy.
Adarsh Shirawalmath, Machine Learning Engineer at Hugging Face, notes that enterprises and people want to stay close to their data and control their own infrastructure, and inference economics make self-hosting appealing.
“We also see it in the numbers on the Hub: open models get downloaded, fine-tuned and adapted at a scale no closed API can match, and that ecosystem effect is hard to ignore,” he says.
Adam Mainz, a software engineer at Google, believes Muse Glimmer’s smaller size makes it more practical for local developers. Another user, replying to his post on X, wrote, “I ran Muse Glimmer on a single RTX 4090, and it was fast enough for a real agentic task, not just a benchmark.”
Shirawalmath says that the success of an open-weight model depends on how easily developers can actually use and build on it. “Real performance per parameter at sizes people can actually run, and day zero support in the tools developers already use like SGLang, Transformers, etc., and a good permissive license” are key factors, he notes.
Meta looks determined to recreate the effect Llama had on the tech community. It gave the company influence far beyond the number of users interacting directly with Meta AI. That's what it wants for Muse Glimmer.