Meta’s local agent bet now has a name: Muse Glimmer, a 30B open-weight multimodal model released on August 10 under Apache 2.0, built to run always-on agent workflows on a single consumer GPU.
What Happened
The announcement arrived on August 10 through three channels at once. Meta’s AI account posted the model card, Mark Zuckerberg published a long essay arguing that everyone should have access to superintelligence, and Scale AI founder Alexandr Wang, who now serves as Meta’s chief AI officer, added that the company will also open the weights of its larger foundation model, Muse Spark 1.2, soon.
Muse Glimmer is a 30B-parameter dense multimodal model with a 128k-plus token context window. Meta designed it for local, always-on agent work: models that sit on your own machine and run continuously rather than being called from a cloud API. The weights are hosted on Hugging Face under the permissive Apache 2.0 license, the same one Meta used for Llama, and research details are published on research.meta.ai.
The model is built to run entirely on consumer hardware. Meta says a Mac or a PC with a capable GPU is enough, and Wang’s thread specifies that a 24GB card runs it without losing agentic reliability. Independent tests on the SGLang side confirm the reach: Muse Glimmer runs on Apple Silicon through the MLX backend, on NVIDIA RTX 5090 and RTX PRO 6000 cards, and on the DGX Spark mini workstation.
What Is Inside
The architecture is a hybrid of image and text components rather than a single uniform transformer. The model stacks a 27.9B dense text decoder with a 1.9B vision transformer and a GELU-based multimodal projector that maps visual features into the text decoder’s space.
The text decoder has 52 transformer layers. Each layer uses grouped-query attention with 32 query heads and two key-value heads, followed by a SwiGLU feed-forward network. The attention pattern is deliberately hybrid: three layers of 2,048-token sliding-window attention are interleaved with a full-sequence attention layer every fourth step. That design matters for context length, because combining rotary position embeddings on the local windows with no-position-embeddings on the full-attention layers lets the model extend well beyond its 128k training window.
Meta Superintelligence Labs partnered with the SGLang team at LMSYS for day-zero inference support, which is unusual for a release of this size. The collaboration produced a dedicated SM120 backend for the latest NVIDIA consumer cards plus an MLX path for Apple Silicon, and both got the same optimizations: DFlash speculative decoding, RadixAttention prefix caching, and breakable CUDA graphs.
The performance numbers from SGLang’s benchmark are the headline. On a single GeForce RTX 5090, the NVFP4-quantized checkpoint with DFlash hits a total output throughput of up to 1,452 tokens per second, with 236 tokens per second of per-user decode speed. That is fast enough for conversational latency on local hardware, which is exactly the use case Meta is aiming at.
The Checkpoints
Meta and SGLang ship several weight formats so the same model fits machines with very different budgets.
- BF16: full fidelity, runs on a single H100 for maximum quality.
- NVFP4: a roughly 19.5GB mixed NVFP4-plus-MXFP8 recipe for the SM120 path.
- NVFP4 + DFlash speculator: an 18GB checkpoint paired with a 5GB BF16 speculator that fits on one RTX 5090.
- GGUF Q4KM: the smaller option, with a group size of 128, for faster inference on tighter memory.
- GGUF Q4K-Dynamic: better quality than Q4KM at similar size.
- MLX 4-bit: the GGUF weights converted for Apple Silicon.
The quantization story is unusually good for a 30B agentic model. Chinese coverage of the release reports that the 4-bit build comes in around 17GB with roughly 1 percent quality loss, which is what makes the 24GB laptop plausible in the first place. Wang’s own framing is that you keep agentic reliability while running the whole stack locally.
How It Compares
Meta claims Muse Glimmer leads its size class on agentic benchmarks. The model scores best on 12 of 24 benchmark suites, beats Google’s Gemma 4 31B on 19 of them, and beats Alibaba’s Qwen 3.6 27B on 14. Its strongest area is agent tasks: on the MCP Atlas benchmark for tool-using agents it scores 75.5 against 62.5 for the closest competitor, a gap that tracks the model’s design purpose.
The honest caveat is that single-model benchmark claims from the releasing lab need a few weeks of independent verification. Llama releases have historically held up reasonably well to third-party re-testing, but the agentic benchmark field is young, and MCP Atlas itself was only published recently. Expect independent evals on agent harnesses like LangGraph or Claude Code-style tool loops before the 75.5 number can be treated as settled.
Why It Matters
The release is a bet on a specific computing future: your agent runs on your machine, not in someone else’s data center. That has privacy and cost implications. A local agent never streams your documents, conversations, or screen contents to a third party, and it costs nothing per token after you buy the hardware. For anyone building agent workflows with sensitive data, that is a genuinely different trade-off from API-based models.
It also puts pressure on the rest of the open-weights world. Gemma 4 31B and Qwen 3.6 27B are the two models Muse Glimmer is directly compared against, and both have strong open-license positions. A three-way race for the best open agentic model at the 30B scale is good for developers, because it keeps the top contenders within reach of a single GPU.
The timing with Spark 1.2 is worth watching. Meta explicitly says the larger foundation model’s weights are coming soon, which suggests Muse Glimmer is the front half of a two-part open-source push: the practical local agent now, the heavier base model shortly after. That would mirror the pattern we saw with the MiniMax-H3 Apple Silicon port, where open weights plus a local runtime turned a cloud-first model into something people actually run on their own desks.
The Limitations
For all the polish, Muse Glimmer is a 30B model trying to do agent work that 70B-plus frontier models still struggle with. Long-horizon tasks, complex multi-tool orchestration, and reliable long-context retrieval remain open problems at this scale. The 128k context window is real but inference at that length on consumer cards will be slow, and the sliding-window attention pattern means effective recall across the full window is not the same as on a dense full-attention model.
There is also the ecosystem question. Apache 2.0 removes legal friction, but local agent stacks are still young: tool schemas, MCP servers, and harness integrations are less mature than the Python and CUDA tooling around traditional LLM inference. Day-zero SGLang support covers the serving side well, but the agent runtime ecosystem around open local models is thinner than what OpenAI and Anthropic offer through their APIs.
The hardware claim deserves scrutiny too. A 24GB card runs the model, but “runs” and “comfortable for always-on agent work” are different things. SGLang’s 1,452 tokens per second figure is impressive on an RTX 5090, yet laptops with 24GB of unified memory share that bandwidth with the whole system. Real-world agent loops, with vision inputs and tool outputs in the context, will land far from the benchmark number.
The Bottom Line
Muse Glimmer is the most credible open-weight local agent model released so far in 2026. The architecture is thoughtful, the quantization story is genuinely good, and the day-zero SGLang and MLX support means you can actually run it this week rather than waiting for the ecosystem to catch up. Treat the benchmark claims as directional until independent evals land, and budget for the reality that agentic workloads at 30B are promising but not frontier-level yet. For developers whose constraint is a single consumer GPU and whose requirement is data staying local, this is the release to test first.