Zuck rekindles open weights Llama drama with Muse Glimmer

After seemingly abandoning its open source AI roots this spring, Meta offered enterprises a glimmer of hope on Monday with the launch of its first open weights model in more than a year. Unveiled on Monday, Muse Glimmer is a 30 billion-parameter LLM distilled from the Social network's larger, and for now proprietary, Muse Spark model. Glimmer arrives as American tech companies grapple with a crisis over the proliferation of Chinese open weights models and a call for domestically-created alternatives. Meta built its reputation on the back of open weights model development beginning with the Llama herd back in 2023. But after Llama 4 flopped and the company restructured its AI group, critics questioned CEO Mark Zuckerberg's commitment to open source AI. With Muse Glimmer, Meta has returned to the open weights arena. The company describes the model as being ideally suited to local AI inference workloads, including local agents, code assistants, and applications requiring robust multi-modal tool use and function calling. Released under a highly permissive Apache 2.0 license, enterprises are also free to deploy, use, and modify the model however they see fit. We imagine it won't be long before Nous Research — one of the original Llama fine-tuners — emits another Hermes model based on Glimmer. Early support has begun to hit popular local AI inference platforms like Llama.cpp, Ollama, and Unsloth, with optimized implementations expected to hit over the next few weeks. But at 30 billion parameters, Muse Glimmer doesn't exactly move the needle much on reclaiming American open weights superiority. It's too small to compete with Moonshot AI's Kimi K3, Alibaba's Qwen 3.8-Max, DeepSeek V4 Flash, or any of the other Chinese models which have dominated the AI news cycle over the past few months. Instead, Muse Glimmer is positioned as a model for small-to-medium sized enterprises or enthusiasts, competing with similarly-sized LLMs from Alibaba and Google. This is reflected in Meta's benchmark figures, which pit the model against Alibaba's Qwen 3.6-27B and Google Gemma 4 31B. As usual, take these claims with a grain of salt, but Glimmer does appear to best Google's Gemma in most scenarios, and trades blows with Alibaba's equivalently sized model. Unfortunately for Meta, the comparison probably won't age well, with Qwen 3.8-27B due to be released any day now. Open and local Glimmer's relatively small parameter count means that its hardware requirements are rather modest compared to larger frontier-class models, like DeepSeek V4. At its native BF16 precision, the model should fit comfortably into a single Nvidia RTX Pro 6000 or AMD MI350P. Quantized to 4-bit precision and model's weights shrink from around 60 GB to just under 16 GB — small enough to fit in a 20 to 24 GB consumer graphics card, like an RTX 30/4090 or RX 7900 XT/XTX. Unfortunately for those with 16 GB cards, a lack of adequate working memory means dropping down to a smaller, less accurate 3-bit quant. Even if you can get the model to fit, it won't be as fast as some other recent models, like Qwen 3.6-35B-A3B or Gemma 4 26B-A4B, as Glimmer uses all 30 billion parameters to generate each token where the others use just 3 to 4 billion. Glimmer benefits heavily from memory bandwidth. On cards like the RTX 5090, which has 1.8 TB/s of it, Meta says users can expect between 75 and 233 tok/s. The higher end of that relies on a technique called speculative decoding. As a quick refresher, speculative decoding uses a small draft model, in this case one modeled after DeepSeek's DSpark drafter, to speed up inference by predicting the outputs of a larger model. If you're interested, we explore the concept in greater depth in this hands-on here. While 233 tok/s is plenty fast for most agentic workloads, most users won't see performance nearly that high. On an M5 Max MacBook Pro, Meta estimates that the systems' up to 614 GB/s of bandwidth will deliver a still-meaningful 26.2 to 57.8 tok/s, but it's also worth pointing out the M5 Max delivers 4x the memory bandwidth of the typical Windows notebook today. In other words, unless you've got a dedicated graphics card with enough memory to run the model, the best you can expect is around 6 to 14 tok/s. Testing on a DGX Spark in Unsloth Studio, we were getting around 12.2 tok/s, though DSpark support doesn't appear to have been implemented just yet. If you'd like to try the model out for yourself, its weights are currently available for download on Hugging Face as well as through local inference platforms like Ollama and LM Studio. Sparking joy Meta's return to the open weights arena doesn't stop with Glimmer. In an X post Monday, Meta's Superintelligence chief Alexandr Wang committed to releasing an open weights version of Muse Spark 1.2 "soon." Muse Spark 1.2 is Meta's most capable model, its first to compete in a meaningful way with rivals OpenAI, Anthropic, and Google. However, even this release won't close the gap with Chinese models. Both Kimi K3 and Qwen 3.8-Max outperform Spark, according to Artificial Analysis' intelligence index. Having said that, we still don't know how big Spark actually is. If Meta's next open weights model ends up delivering similar performance using fewer weights, this may end up being more competitive from a cost per token standpoint. ®