Quick answer. Muse Glimmer is Meta's 30B open-weights model, released August 10, 2026 under Apache 2.0. It has a 128K context, takes text and images, and runs in under 20GB at 4-bit — a single consumer GPU or Mac. It is a tool-calling specialist: it beats Gemma 4 comfortably but loses to Qwen 3.6 on four of seven benchmarks Meta published itself.
Five days after shipping a closed model and a closed coding agent, Meta released a fully open one. Muse Glimmer landed on August 10, 2026: 30 billion parameters, Apache 2.0, ungated weights on Hugging Face, and built to run on hardware you already own.
It went straight to the top of Hacker News — 1,199 points and 637 comments — and picked up roughly 750,000 downloads across the official and community repositories within days. This guide covers what it is, whether it is actually good, and how it compares to the models it is competing with.
What is Muse Glimmer?
A 30-billion-parameter dense, decoder-only multimodal transformer with a built-in vision encoder, from Meta Superintelligence Labs.
- License: Apache 2.0 — genuinely permissive, commercial use included
- Context: 128K tokens (not the 1M of Muse Spark 1.2)
- Input: text and images. Output: text
- Languages: trained on 100+
- Weights: ungated on Hugging Face under the
meta-modelsorganisation
The training recipe is the interesting part. Glimmer was distilled from Muse Spark using logit distillation, then given mid-training on agent-heavy long-context data, then SFT, on-policy distillation and reinforcement learning. It is explicitly a small model taught by a large one, and it was tuned for agentic work rather than for chat.
Can you actually run it locally?
Yes, and this is the headline. Meta engineered it for a specific hardware envelope rather than for a benchmark chart.
| Precision | Memory needed | Practical target |
|---|---|---|
| Full precision | >55 GB | Datacentre / multi-GPU |
| ~4-bit quantized | Under 20 GB | Single 24GB GPU or a 32GB Mac |
That puts it inside reach of an RTX 4090 or 5090, or an Apple Silicon Mac with 32GB of unified memory — the same envelope our Apple Silicon LLM guide and self-hosting guide cover.
It also ships a DFlash speculative-decoding drafter, which is a real practical win rather than a spec-sheet item:
| Hardware | Decode speedup |
|---|---|
| RTX 5090 | 3.1x |
| M5 Max | 1.8x |
| M4 Max | 1.5x |
Runtime support was there on day one — Ollama, LM Studio, llama.cpp, vLLM, SGLang and Unsloth were all launch partners, along with AMD, Arm, Dell, Hugging Face, Intel, NVIDIA and Together AI. Community quantizations appeared within about 24 hours across MLX, ROCm, NVFP4, FP8 and abliterated variants.
If you would rather not host it, it is on OpenRouter as meta/muse-glimmer-30b at $0.35 input and $1.50 output per million tokens.
Is Muse Glimmer any good?
It depends entirely on what you want it for — and Meta's own benchmark table is unusually honest about this.
Meta compared Glimmer against its actual size class, Gemma 4 31B and Qwen 3.6 27B, both in thinking mode:
| Benchmark | Muse Glimmer | Gemma 4 31B | Qwen 3.6 27B |
|---|---|---|---|
| SWE-Bench Verified | 76.0 | 66.6 | 77.2 |
| SWE-Bench Pro | 51.2 | 36.9 | 50.2 |
| TerminalBench 2.1 | 51.7 | 43.4 | 60.7 |
| OSWorld-Verified | 65.9 | 58.5 | 75.6 |
| MCP Atlas | 75.5 | 54.2 | 62.5 |
| WildClawBench | 47.6 | 37.6 | 43.2 |
| SkillsBench (with skills) | 44.3 | 32.4 | 46.6 |
Qwen 3.6 27B beats Muse Glimmer on four of seven rows — including SWE-Bench Verified, TerminalBench 2.1 and OSWorld. Publishing that is a creditable choice, and it tells you exactly how to read the model.
Glimmer's three wins share a shape: MCP Atlas (+13 over Qwen), WildClawBench and SWE-Bench Pro are all tool-calling and agentic-orchestration benchmarks. The Hacker News consensus put it bluntly — it "barely edges out Qwen3.6 27B except for tool-calling skills."
So the accurate summary is: Muse Glimmer is a tool-use specialist in its size class, not a general coding leader. If you are building agents that call MCP servers and chain tools, the MCP Atlas gap is substantial and it is the most interesting open model in this weight class. If you want the best local coding model outright, Qwen 3.6 is still ahead on the benchmarks Meta chose.
Against Gemma 4 31B, Glimmer wins every single row — that comparison is not close.
What are the catches?
It is dense, not mixture-of-experts. This was the loudest technical criticism at launch. A dense 30B activates all 30 billion parameters per token, which makes it memory-bandwidth-bound and therefore slower than same-size MoE competitors that activate a fraction of their weights. The DFlash drafter offsets some of this, but the architecture choice is real and it shows up as tokens per second on consumer hardware.
128K context, not 1M. A big step down from Muse Spark 1.2's million-token window. Fine for most work, limiting for whole-monorepo tasks.
Meta does not host it. There is no first-party API and no rate-limited free tier, which surprised a lot of people. You self-host or use a third party.
Why did Meta open this one?
Because the strategy reversed, and Glimmer is only half of it.
Through the first half of 2026 the story was that Meta had abandoned open weights: Llama gave way to the closed Muse Spark family in April, and Muse Code shipped on August 5 as a closed, paid product.
Then on August 10 Meta released Glimmer under Apache 2.0 and Mark Zuckerberg announced that Muse Spark 1.2's own weights will be open-sourced, publishing a roughly 6,500-word essay arguing for American open-source AI. CNBC, the Financial Times and Fortune all covered it.
One important caveat: the Muse Spark 1.2 weights are not out yet. As of mid-August 2026 the meta-models Hugging Face organisation contains only the Glimmer repositories, and no release date has been given. Treat Spark 1.2 open weights as announced, not shipped.
The community reaction split roughly three ways: genuine relief that Meta was back to open weights, scepticism that Glimmer only narrowly beats Qwen outside tool-calling, and irritation at the dense-versus-MoE choice.
Should you use Muse Glimmer?
Yes, if you are building local agents that lean on tool calling and MCP. The MCP Atlas result is the best in its weight class by a wide margin, the Apache 2.0 license imposes no conditions, and it fits a 24GB card.
Yes, if you want a capable multimodal local model and value the permissive license. Gemma 4 31B is beaten on every benchmark Meta published.
Probably not, if you want the strongest local coding model — Qwen 3.6 27B still leads on SWE-Bench Verified, TerminalBench and OSWorld.
Probably not, if tokens per second is your binding constraint. A dense 30B will be slower than an MoE of similar size on the same hardware.
For where it sits in the wider field, see our open-source LLM landscape guide.
FAQ
What is Muse Glimmer?
Meta's 30B open-weights multimodal model, released August 10, 2026 under Apache 2.0. It has a 128K context window, accepts text and images, and was distilled from Muse Spark and tuned for agentic tool use.
Is Muse Glimmer really open source?
The weights are released under Apache 2.0 and are ungated on Hugging Face, so commercial use and fine-tuning are permitted. As with most "open" model releases, the training data is not published.
What hardware do I need to run Muse Glimmer?
At roughly 4-bit quantization the language model fits under 20GB, targeting a 24GB GPU or a 32GB Mac. Full precision needs more than 55GB.
Is Muse Glimmer better than Qwen 3.6?
Not overall. On Meta's own table Qwen 3.6 27B wins four of seven benchmarks including SWE-Bench Verified and TerminalBench 2.1. Glimmer wins the tool-calling benchmarks — notably MCP Atlas, 75.5 against 62.5.
Is Muse Glimmer better than Gemma 4?
Yes, on every benchmark Meta published, several by wide margins.
How fast is Muse Glimmer?
It ships a DFlash speculative-decoding drafter giving 3.1x decode speedup on an RTX 5090, 1.8x on an M5 Max and 1.5x on an M4 Max. Being dense rather than MoE, it is memory-bandwidth-bound and slower than comparable MoE models.
Can I run Muse Glimmer in Ollama or LM Studio?
Yes. Ollama, LM Studio, llama.cpp, vLLM, SGLang and Unsloth were all launch partners, and GGUF builds were available immediately.
Is Muse Spark 1.2 open source too?
Not yet. Zuckerberg announced on August 10, 2026 that its weights will be open-sourced, but they have not been released and no date has been given.
What is the difference between Muse Glimmer and Muse Spark 1.2?
Glimmer is 30B, open-weights, 128K context and runs locally. Muse Spark 1.2 is a much larger closed model with a 1M context, served only through Meta's API at $1.25/$4.25 per million tokens, and it powers the Muse Code agent.