Guide
OpenClaw + Muse Glimmer: Meta's Open 30B Agent Model
Meta released Muse Glimmer on 10 August 2026 — a 30B dense model under Apache 2.0, distilled from the closed Muse Spark and built to run long agent sessions on a single consumer GPU. It is the first Meta agent model you can actually download. Here is what it needs, what it costs hosted, and how to point an OpenClaw bot at it.
Glimmer Is Not Muse Spark
These two ship under the same brand and get confused constantly, so start here. Muse Spark 1.2 is proprietary, has a 1M-token context, and is reachable only through an API. Muse Glimmer is the open release: 30B parameters, weights on Hugging Face, Apache 2.0, and a 128K context. Meta describes Glimmer as distilled from Spark — the same family, shrunk to fit on hardware you own.
The practical difference: Spark is what you rent, Glimmer is what you keep. If your reason for reading this is “I want an agent model with no vendor in the loop,” Glimmer is the one.
Specs
| Property | Value |
|---|---|
| Hugging Face repo | meta-models/Muse-Glimmer-30B |
| Parameters | ~29.6B total, including the vision encoder |
| Architecture | Dense causal transformer — not mixture-of-experts |
| Context window | 131,072 tokens |
| Modalities | Text + image in → text out (max 4,096 visual tokens per image) |
| Licence | Apache 2.0 |
| Languages | Trained across 100+ |
| OpenRouter ID | meta/muse-glimmer-30b — $0.30 / 1M in, $1.20 / 1M out |
Dense rather than MoE is a deliberate choice for agent work. Every parameter fires on every token, so latency is predictable and there is no routing layer to behave oddly on the two-hundredth step of a long task. It costs more memory per unit of capability than a sparse model, and buys back consistency.
What It Scores
Meta published Glimmer at 76.0 on SWE-Bench Verified and 51.2 on SWE-Bench Pro for coding, 75.5 on MCP Atlas and 74.6 on DeepSearch QA for general agentic work, plus 78.8 on Charxiv Reasoning and 75.8 on OmniDocBench on the multimodal side.
These are vendor-published numbers and should be read as such. The one worth weighting is MCP Atlas: it measures tool-calling across an MCP server surface, which is close to what an OpenClaw bot spends its day doing. A 30B model landing in that range is the actual news — not the SWE-Bench line.
The VRAM Question
“Runs on a single GPU” is doing some work in the press coverage. Which single GPU depends entirely on how hard you quantize:
| Build | VRAM needed | Realistic hardware |
|---|---|---|
| Full precision (BF16) | 64 GB | Workstation card, or unified memory — see Strix Halo |
| K-Quant-Dynamic | 32 GB | RTX 5090, A6000, 32GB+ Apple Silicon |
| K-Quant-17GB (4-bit) + DFlash | 24 GB | RTX 4090, RTX 3090 |
DFlash is a companion drafter head Meta ships for speculative decoding — it guesses several tokens ahead and lets the main model verify them in one pass, which is where the 24GB tier gets its speed back. There are already 96+ community quantizations for llama.cpp, Ollama, LM Studio and Jan, so the menu is wider than the official three.
Route 1: Hosted Bot via OpenRouter
If you want a Glimmer agent answering messages around the clock without leaving a GPU running at home, rent the inference and host the agent. At $0.30 in and $1.20 out per million tokens, Glimmer is roughly a quarter of Muse Spark's input price.
- Create a key at openrouter.ai/keys. No minimum spend.
- Add it on your API Keys page, or paste it into the configurator under BYOK — the BYOK guide has the full walkthrough.
- Set the model to
meta/muse-glimmer-30b. - Pick a channel — Telegram, Discord, WhatsApp, WeChat or the web gateway — and click Deploy.
Already have a bot running? Switch without redeploying by sending /model meta/muse-glimmer-30b from any connected chat. Check the model list for what is available as a one-click pick today.
Route 2: Run the Weights Yourself
The Ollama path is one command:
ollama pull muse-glimmer:30b # 18 GB
ollama pull muse-glimmer:30b-mlx # 21 GB, Apple Silicon + DFlashThe 30b-mlx tag is built for Apple Silicon and keeps image input working. For a serving setup rather than a laptop setup, vLLM exposes Glimmer over an OpenAI-compatible API — see OpenClaw + vLLM. llama.cpp handles the mixed CPU/GPU case if you are short on VRAM.
localhost is not routable from someone else's machine. To pair a hosted agent with your own Glimmer, the endpoint needs a public HTTPS address (a tunnel, or a box with a domain). Otherwise, run both halves locally: see OpenClaw + Ollama.Self-Hosted Configuration
Running your own OpenClaw against a local Ollama serving Glimmer, this is the whole change in openclaw.json:
{
"models": {
"providers": {
"ollama": {
"baseURL": "http://localhost:11434/v1"
}
}
},
"agents": {
"defaults": {
"model": {
"primary": "ollama/muse-glimmer:30b"
}
}
}
}For the rented route instead, swap the provider block for OpenRouter and use openrouter/meta/muse-glimmer-30b. Full provider setup lives in the OpenRouter guide.
Is It Worth Switching?
Honestly: it depends on what your agent does. For long-running tool use on hardware you control, Glimmer is the strongest open option Meta has put out, and Apache 2.0 means no licence to argue about. For hardest-case reasoning it is still a 30B model, and the frontier closed models are ahead. A sensible setup keeps Glimmer as the everyday driver and a larger model configured as the escalation path — see best models for OpenClaw.
FAQ
Can I run Muse Glimmer on an RTX 4090?
Yes. The 4-bit K-Quant-17GB build with the DFlash drafter fits in 24GB, which is exactly a 4090 or a 3090. Full BF16 precision needs 64GB and will not fit.
Is Muse Glimmer free?
The weights are, under Apache 2.0 — download and run them commercially with no fee. Renting it through an API provider such as OpenRouter costs $0.30 per million input tokens and $1.20 per million output tokens.
Does Muse Glimmer understand images?
Yes. It takes text and images in and returns text, with a frozen ViT-G/14 perception encoder handling vision at up to 4,096 visual tokens per image. It does not generate images.
How is Glimmer different from Muse Spark 1.2?
Glimmer is open-weight, 30B, 128K context, and self-hostable. Spark 1.2 is proprietary, much larger, has a 1M-token context, and is API-only. Glimmer is distilled from Spark.