Meta just made its strongest open weights move in years. Today the company released Muse Glimmer, a 30-billion-parameter agentic model built to run on a single consumer GPU, and open-sourced it under Apache 2.0. Then Zuckerberg went on X and said the weights for Muse Spark 1.2, Meta’s actual frontier model, are coming too.
That second part is the big one. Muse Spark 1.2 isn’t a toy. It’s the model that powers Muse Code, Meta’s terminal coding agent, and it’s currently the fifth most powerful model on the Artificial Analysis Intelligence Index. If Meta ships those weights, it becomes the most capable open model from an American lab in years.
Why these open weights matter
Glimmer is the here-and-now piece. It’s a 30-billion-parameter dense model optimized for always-on local agent workflows. Meta says it’s small enough to run on a Mac or PC with a single consumer GPU, and it’s built for local agents, function calling, local coding, and LLM-as-a-judge evaluation.
The training story is interesting too. Meta trained Glimmer on Muse Spark’s outputs using logit distillation, then layered in longer-context, agent-heavy data with richer reasoning traces, then fine-tuned with a mix of on-policy distillation and reinforcement learning.
Here’s the part that matters for anyone with a real GPU. At full precision, a 30-billion-parameter model needs over 55 GB of memory, which no consumer GPU has. Meta uses quantization to compress the weights to roughly 4-bit precision, shrinking the language model to under 20 GB. That leaves room for the KV cache, the perception encoder, and the speculative decoding drafter inside a 24 GB or 32 GB envelope.
The drafter is the clever bit. It’s a small companion model based on DFlash that proposes whole blocks of tokens at once while the main model verifies them in parallel. Meta measured the quantized model on a MacBook M4-Max, M5-Max, and an RTX-5090, and says it’s fast enough for fluid conversation and real-time agent interaction, entirely on-device.
What Glimmer actually does
This isn’t a chat model with a local sticker. Meta’s pitch is that Glimmer is a real agent, not a chatbot wearing a local-mode costume. The company says it achieves strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, tau-Bench and SWE-Bench, which test whether a model can work inside scaffolds, write and debug code, and resolve multi-turn requests from start to finish. When a tool call fails or returns something unexpected, Glimmer is trained to diagnose the error and retry rather than halt.
It accepts interleaved text and images through a dedicated perception encoder, so an agent can read screenshots, charts, and documents. It works across OpenClaw and other agentic orchestration patterns, it supports different reasoning strengths so you can trade quality for speed, and it’s trained on data from more than 100 languages.
Meta is positioning it against Google’s Gemma family and the Chinese local models. The benchmarks compare Glimmer with Gemma4-31B and Qwen3.6-27B, and Meta says it performs strongly for its size class. That’s the right fight. Gemma has spent a year building momentum in the on-device space, and Alibaba’s Qwen and DeepSeek have owned the efficient local model category.
Where you can actually run it
The integrations are the part that makes this real for normal users. Meta says that in the coming days you can run Glimmer locally through Ollama, LM Studio, and Unsloth, deploy it with edge frameworks like llama.cpp, ExecuTorch, and MLX, or serve it at scale with vLLM and SGLang. There are also hosted paths through Together AI, Fireworks AI, and OpenRouter, plus a PyTorch TorchTitan route for customizing the model.
Meta says it’s working with AMD, Arm, Dell, Intel, and NVIDIA to optimize performance across devices, and the weights are already on Hugging Face. That’s the full open-source loop: download today, tune it to your own workflows, run it on your own hardware, keep your data local.
I’ve been tracking this open-weights wave all year, and the pattern is consistent. The local model race is where the open weights action actually lives, because that’s where developers can self-host without burning cloud credits. Meta is late to this particular party, but it showed up with a genuinely agentic model rather than another chat-tuned small model.
The Spark 1.2 promise
Here’s what I can’t stop thinking about. Zuckerberg said the Muse Spark 1.2 weights are coming “soon,” which is not a date. The best AI coding agents of 2026 all run on closed frontier models right now. If Meta actually ships Spark 1.2 weights, the entire calculus changes, because you’d be able to run a frontier-adjacent coding model on your own hardware.
Glimmer is the open weights release that ships today. Spark 1.2 is the promise that reshapes the market if it lands. Meta has been burned on open source before, with Llama 4’s misleading benchmark numbers and the slow walk back from open commitments. So I’ll believe the Spark weights when they’re on Hugging Face.
Still, the direction is clear. Meta is fighting on two fronts at once, frontier open weights with Spark 1.2 and local-first open weights with Glimmer. For developers who want agents they can actually run and own, today is a genuinely good day. And if the Spark 1.2 weights actually land, the open model conversation changes completely.
