Guide
Hermes Agent + Muse Glimmer: Meta's Open 30B Model
Muse Glimmer landed on 10 August 2026 — Meta's first downloadable agent model, 30B dense parameters under Apache 2.0 with a 128K context. Hermes Agent takes any OpenAI-compatible endpoint, which means Glimmer works both ways: rented through a provider, or served off your own GPU with nothing leaving the building.
Why This Pairing Makes Sense
Hermes is the framework people pick when they want the agent on their own terms — own box, own data, own rules. Until now the strongest agent models were all API-only, so “self-hosted Hermes” still meant shipping every prompt to someone else's inference. An Apache 2.0 model that tool-calls competently and fits on one GPU closes that gap.
Glimmer is distilled from Muse Spark 1.2, Meta's closed flagship. Spark keeps the 1M-token context and the higher ceiling; Glimmer trades those for weights you can actually hold.
Specs at a Glance
| Property | Value |
|---|---|
| Weights | meta-models/Muse-Glimmer-30B on Hugging Face |
| Parameters | ~29.6B dense, vision encoder included |
| Context | 131,072 tokens |
| Licence | Apache 2.0 — commercial use allowed |
| Modalities | Text + image in → text out |
| Agentic benchmarks | MCP Atlas 75.5, DeepSearch QA 74.6, SWE-Bench Verified 76.0 |
| OpenRouter ID | meta/muse-glimmer-30b — $0.30 / 1M in, $1.20 / 1M out |
MCP Atlas is the number to care about here. It scores tool use across an MCP server surface, and a Hermes agent lives or dies on whether the model emits clean tool calls on step forty of a task rather than narrating what it would do. Benchmarks are vendor-published, so treat them as a starting hypothesis, not a guarantee.
Option 1: Hosted Hermes on OpenClaw Launch
The fastest route — no GPU, no serving stack. Glimmer is rented through OpenRouter and your Hermes agent runs on our infrastructure:
- Get a key at openrouter.ai/keys.
- Add it on the API Keys page — see the Hermes BYOK guide for the walkthrough.
- Choose Hermes Agent in the configurator and set the model to
meta/muse-glimmer-30b. - Pick your channel and deploy.
Switch at runtime from any connected channel with /model meta/muse-glimmer-30b — no redeploy.
Option 2: Self-Hosted Hermes, Rented Inference
Running your own Hermes but happy to call an API for the model:
export OPENROUTER_API_KEY=sk-or-...
hermes inference set openrouter
hermes model set meta/muse-glimmer-30bOr edit the config file directly:
# /opt/data/config.yaml
inference:
provider: openrouter
model:
default: meta/muse-glimmer-30bFull provider setup lives in the Hermes + OpenRouter guide.
Option 3: Fully Local — Nothing Leaves the Box
Pull the weights, serve them, point Hermes at the local endpoint.
ollama pull muse-glimmer:30b # 18 GB
ollama serve # OpenAI-compatible on :11434/v1Hermes accepts any OpenAI-compatible base URL as its provider endpoint. Point it at http://localhost:11434/v1 when Hermes shares the host with Ollama, or http://host.docker.internal:11434/v1 when Hermes runs in Docker on Mac or Windows.
The exact config field names move between Hermes versions, so confirm them against the upstream README for the version you actually have installed — look for the option that sets a custom OpenAI API base URL. Hermes does not hot-reload config; restart the container after saving.
For a throughput-oriented setup rather than a laptop one, vLLM serves Glimmer over the same interface — see Hermes Agent + vLLM. The broader local setup is covered in Hermes Agent + Ollama.
What Hardware You Actually Need
| Build | VRAM | Notes |
|---|---|---|
| BF16 full precision | 64 GB | Best quality; needs a workstation card or unified memory |
| K-Quant-Dynamic | 32 GB | The sensible default for most self-hosters |
| K-Quant-17GB (4-bit) + DFlash | 24 GB | RTX 4090 / 3090; DFlash speculative decoding recovers speed |
A 128GB unified-memory box runs the full-precision build with room left for context — see AMD Strix Halo for what that costs in 2026. Check Hermes system requirements for the agent side, which is modest next to the model.
localhost — that address means something different on every machine. Either self-host both halves, or expose your model server on a public HTTPS address before wiring it up.FAQ
Can Hermes Agent run Muse Glimmer without any API key?
Yes, if you serve the weights yourself. Ollama, llama.cpp or vLLM expose an OpenAI-compatible endpoint locally and no key or external call is involved. An API key is only needed for the rented route through a provider.
Which is better for Hermes, Glimmer or Muse Spark 1.2?
Spark has the higher ceiling and a 1M-token context, so it wins on the hardest long-context reasoning. Glimmer wins whenever self-hosting, data control, or cost per token matters more than peak capability — and it is about a quarter of Spark's input price when rented.
Does Muse Glimmer handle tool calling reliably?
Meta tuned it specifically for tool use, long tasks and failure recovery, and reports 75.5 on MCP Atlas. That is strong for a 30B model, though vendor-published. Test it against your own tool set before trusting it with unattended work.