Artificial Intelligence & Machine Learning
,
Next-Generation Technologies & Secure Development
Local Agentic AI Model Targets Coding, Tool Calling and Multi-Step Tasks
Emilia David •
August 10, 2026
Mark Zuckerberg speaks at Georgetown University in Washington, D.C., in a photo dated Sept. 19, 2025. (Image: FotoField/Shutterstock)
Meta hopes to recapture the momentum it had when it first launched its Llama artificial intelligence model. Now, with a new model and an increased focus on open-source AI, the social media giant is going against the more proprietary approach of its competitors.
See Also: How AI Agents Widen the Enterprise Blast Radius
The new model, dubbed Muse Glimmer, is a 30-billion-parameter model that can run locally on a single GPU or computer. Unlike other AI models made by U.S.-based frontier labs, Muse Glimmer leans on the open-source ethos, running on an Apache 2.0 license. By focusing on its openness, Muse Glimmer hews closely to Meta CEO Mark Zuckerberg’s newly articulated AI philosophy that open-source models are a way for American models to compete against Chinese labs.
The new model does not aim to match Mythos 5 or other cutting edge frontier models. It’s a smaller model that can code and deliver agentic tasks. LLMs like Mythos 5 mainly run on cloud instances in outside data centers because they need server clusters to operate. Users make API calls when prompting a model like that. Muse Glimmer runs locally, meaning it does not depend on cloud infrastructure or internet access.
“This is increasingly viable: the open-source community has shown that smaller models, when trained effectively, can approach frontier-level performance on targeted tasks,” Meta said in its blog post.
Meta’s released frontier model Llama 1 to fanfare in 2023, with the social media company touting a relatively more permissive open-source license. But as enterprises used the model to start building applications, Meta drew criticism for burying the fact that the model’s openness came with a caveat around commercialization, suggesting that it was not truly operating under the open-source ethos.
The company continued to release Llama models, but the release of powerful reasoning models from OpenAI, Anthropic and Google, along with the entrance of Chinese open-weight models, eroded Meta’s leaderboard rankings.
Muse Glimmer works best for always-on, local, agentic use cases such as coding, tool calling and multi-step tasks without depending on an internet connection. Much like OpenClaw agents, people can use Muse Glimmer for workflows that run within their computers.
Although they’re hardly the stuff of breathless PR announcements, frontier labs do offer smaller models that can run locally, such as OpenAI’s open-sourced gpt-oss and Google’s Gemma family of models.
Based on Meta’s benchmarks, Muse Glimmer outperforms Gemma 4-31B and Qwen 3.6-27B in some tests. For agentic coding, Muse Glimmer scored 51.2 on the SWE-Bench Pro evaluation compared to 36.9 for Gemma 4 and Qwen’s 50.2. In other coding benchmark tests, like TerminalBench 2.1, Qwen scored significantly higher.
Distillation as a Competitive Stance
Muse Glimmer’s design and stated purpose – to be a model that more people can use – aligns with a 6,500-word manifesto from Zuckerberg, in which the social media titan said he believes in open-source technology, which puts “power into people’s hands.”
“If the power of superintelligence is held by a small number of individuals, businesses, governments, or AI itself, then that will naturally lead to outcomes that are less favorable for everyone else,” Zuckerberg wrote.
Zuckerberg also defended distillation training, a practice that lets AI developers train a model using the responses of another “teacher model” so they don’t need to bring in large datasets to learn from. Anthropic and OpenAI asked the government for help to stop frontier labs, mainly in China, from distilling their models to train competitive LLMs without permission.
He said that for the United States to become a leader in open-source AI, “we will need to rethink our policies in several areas,” including rejecting the notion that distillation is harmful and not restricting access to foreign open-source models.
Zuckerberg leaned heavily into the idea that frontier and smaller models should be in service of creating personal agents for individuals.
“Most other labs are focused on building AI for companies, governments or other institutions, so if those labs lead, then the balance of power will favor larger institutions over individuals,” Zuckerberg said.
These comments provide a bit of a turnaround from Zuckerberg’s previous comments. As reported by TechCrunch, the Meta CEO said during the company’s second quarter earnings call that he sees a strong opportunity around the enterprise segment to sell its business agents, APIs and compute.