A 30-billion-parameter model built specifically for always-on local agents, fitting inside a 24GB VRAM envelope, shipping under Apache 2.0 - that combination did not exist three weeks ago. Meta shipped its first open-weight model built specifically for always-on local agent work on August 10, 2026. Muse Glimmer is a 30B dense multimodal model released under Apache 2.0, distilled from the much larger Muse Spark family, and tuned for the things agents actually do: calling tools, recovering from failures, reading screenshots, and holding 131K-token contexts.
The conversation around local models usually starts with "can I run it at all?" Glimmer shifts that to "what does it actually do once it's running?" That is a more interesting question.
What Muse Glimmer is, exactly
Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows. It is small enough to run on a Mac or PC with a single consumer GPU, enabling use cases that range from local agents and function calling, to local coding, and LLM-as-a-judge evaluation.
Meta's Superintelligence Lab published Muse Glimmer in GGUF form under Apache 2.0 - a ~29.6B multimodal model distilled from a larger Muse Spark.
Image input is handled by a roughly 1.8B ViT-G/14 perception encoder shipped as a separate mmproj file, though the model is text-out only and does not treat audio or video as primary modalities.
The distillation lineage matters. Training a 30B model for that balance took three phases: pre-training on Muse Spark's outputs using logit distillation with a similar data mix to the teacher model; mid-training on longer-context, agent-heavy data with richer reasoning traces; and post-training using supervised fine-tuning combined with on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains.
The upshot: this is not a repurposed chat model with tool-call examples bolted on. The agent capability was the training objective from phase two onward.
The hardware numbers - and what they mean for agent loops
Muse Glimmer accepts text and images, has a stated context length of 131,072+ tokens, supports more than 100 training languages, and is released under Apache 2.0 for commercial and research use. Meta targets 24GB hardware with a 17GB 4-bit quant and 32GB hardware with its less compressed dynamic quant, including headroom for the vision encoder, KV cache, and DFlash drafter.
The speed figure deserves a closer look. A companion DFlash drafter for speculative decoding is bundled alongside the weights, and Meta says it pushes an RTX 5090 from 74.9 to 233.4 tokens per second - a 3.1× speedup. That matters differently for agents than for chat. A chat session might run 200 tokens of output. An agent loop calling five tools, writing a file, verifying a result, and summarizing might issue 2,000-4,000 tokens across a single task. At 74.9 tok/s that is 27-53 seconds of generation time; at 233.4 tok/s it drops to 9-17 seconds. Latency compounds in agent loops in a way it does not in single-turn use.
What the benchmarks actually show
In Meta's published results, Glimmer leads Gemma4-31B and Qwen3.6-27B on MCP-Atlas, DeepSearch QA, and SWE-Bench Pro, but trails Qwen on Terminal-Bench 2.1, OSWorld-Verified, and several multimodal tests.
| Benchmark | Muse Glimmer 30B | Qwen3.6-27B | Winner |
|---|---|---|---|
| MCP Atlas (tool orchestration) | 75.5 | 62.5 | Glimmer |
| DeepSearch QA | 74.6 | 71.1 | Glimmer |
| SWE-Bench Pro | 51.2 | 50.2 | Glimmer |
| AIME 2026 (math reasoning) | 94.7 | 94.1 | Glimmer |
| OSWorld-Verified (OS tasks) | 65.9 | 75.6 | Qwen3.6 |
| Terminal-Bench 2.1 | 51.7 | 60.7 | Qwen3.6 |
| SWE-Bench Verified | 76.0 | 77.2 | Qwen3.6 |
All figures Meta-reported; no independent third-party replication as of August 26, 2026.
The pattern is meaningful: Glimmer wins on tool orchestration and document reasoning, loses on terminal and OS-level task completion. If your agent spends most of its time calling structured APIs and reading files, the benchmark profile favors Glimmer. If it is running shell commands or driving a desktop GUI, Qwen3.6 has the edge - and Qwen3.8-27B, which shipped roughly the same week, has not been benchmarked against Glimmer yet.
The comparison is against the four-month-old Qwen3.6; Qwen3.8-27B was expected the same week. That is the single most important context the headline numbers omit.
Scaffold compatibility and the "open-weight" caveat
Muse Glimmer works across OpenClaw, Hermes Agent, and other agentic orchestration patterns. That is a genuinely useful detail - it means teams already running an open-source agent framework do not need to rewire their scaffold to pilot this model. For a team evaluating local inference for a coding agent workflow, the setup path is already documented.
On the license question: Meta publishes the BF16 weights, two official GGUFs, the vision projector, and the DFlash drafter under Apache 2.0 - unusually permissive for commercial modification and redistribution. But the release does not include the complete training dataset, every data-cleaning decision, or a reproducible training pipeline. Meta's model card also links a separate Usage Policy covering prohibited uses and deployment responsibilities. "Open-weight" is therefore the precise term.
Apache 2.0 on the weights is better than most comparable releases. It is not the same as open-source. If your legal team's question is "can we fine-tune and redistribute this commercially," the answer is yes. If the question is "can we reproduce the training run," the answer is no.
Beagle in action#ai-infra, mid-sprintA local model running at 233 tokens per second on a single GPU is fast enough to be genuinely useful in an agent loop. Whether Muse Glimmer is the right model for your loop depends on what that loop actually does. The MCP Atlas lead is the clearest signal: if the task is structured tool use, this is a serious option. If it is terminal work or GUI automation, wait for the Qwen3.8 comparison.
Muse Glimmer 30B local agent: common questions
What hardware do I need to run Muse Glimmer 30B?
Meta targets 24GB hardware with a 17GB 4-bit quant and 32GB hardware with its less compressed dynamic quant, including headroom for the vision encoder, KV cache, and DFlash drafter. An RTX 4090, RTX 5080, or Mac with 32GB unified memory are all viable. The 24GB configuration is tight; if you run other GPU workloads concurrently, target 32GB.
Is Muse Glimmer actually open source?
Muse Glimmer is open-weight under Apache 2.0. The release does not include the complete training dataset, every data-cleaning decision, or a reproducible training pipeline. Meta's model card links a separate Usage Policy covering prohibited uses and deployment responsibilities. "Open-weight" is therefore the precise term. Commercial use and fine-tuning are permitted; reproducing the training run is not possible from the released artifacts.
How does Muse Glimmer compare to Qwen3.6-27B for coding agents?
Glimmer leads on tool orchestration (MCP Atlas: 75.5 vs 62.5) and SWE-Bench Pro (51.2 vs 50.2), but Qwen3.6 leads on OSWorld-Verified (75.6 vs 65.9) and Terminal-Bench 2.1 (60.7 vs 51.7). For API-calling and document-reasoning agents, Glimmer has the edge. For terminal-heavy or OS-level automation, Qwen3.6 does. And Qwen3.8-27B has not yet been benchmarked head-to-head.
What is DFlash and do I need it?
DFlash is Meta's speculative decoding drafter, shipped alongside the Glimmer weights. Meta says it pushes an RTX 5090 from 74.9 to 233.4 tokens per second - a 3.1× speedup. For a chat use case the difference is noticeable. For an agent loop that issues dozens of sequential generations per task, it is the difference between a usable and an unusable experience. Enable it.
Which agent frameworks work with Muse Glimmer?
The model card names OpenClaw and Hermes Agent as tested orchestration patterns, alongside deployment through llama.cpp, vLLM, Ollama, LM Studio, and Jan. If you are already running one of those, you can drop Glimmer in without rewriting your scaffold.