Guide
Muse Spark 1.3 on OpenClaw and Hermes
Meta released Muse Spark 1.3 on 2 September 2026, four weeks after 1.2. The headline change is a max reasoning mode built for agentic tasks — which is exactly the workload an always-on agent generates. Here is what changed, how to point an OpenClaw or Hermes agent at it, and when the extra thinking is worth what it costs.
What Changed in 1.3
Muse Spark has been Meta's agent-oriented reasoning line since April 2026. Version 1.2 established the shape most people know it by: a very large context window and pricing well under most models in its class. Version 1.3 keeps that and adds a max reasoning setting targeted at agentic work — longer internal deliberation before the model commits to a tool call or an answer.
That is a meaningful distinction for agents specifically. A chat model that reasons harder produces a better paragraph. An agent model that reasons harder makes fewer wrong tool calls, and a wrong tool call is expensive in a way a mediocre paragraph is not — it costs a round trip, sometimes a side effect you have to undo.
Release Timeline
| Version | Released | Notes |
|---|---|---|
| Muse Spark | 8 April 2026 | Initial release of Meta’s agent-focused reasoning line. |
| Muse Spark 1.1 | 9 July 2026 | Iteration on the original. |
| Muse Spark 1.2 | 5 August 2026 | 1M-token context, aggressive pricing, and the debut of Muse Code. |
| Muse Spark 1.3 | 2 September 2026 | Adds a max reasoning mode aimed at agentic tasks. |
How to Run It on OpenClaw Launch
- Deploy an agent from openclawlaunch.com — about 30 seconds. Bring-your-own-key model selection needs a paid plan; the free trial runs a fixed set of free models.
- In your dashboard, add your OpenRouter key under BYOK settings, or a direct Meta Model API key if you have one.
- Open the model picker and type muse spark. The picker searches the live provider catalog, so 1.3 appears as soon as your provider lists it — no redeploy, no config file edit.
- Select it. The change applies to the running agent immediately.
Self-Hosted: OpenClaw
OpenClaw takes a single string that names the route it goes through, so an OpenRouter setup carries the aggregator as the first segment and the vendor slug after it:
agents:
defaults:
model:
primary: "openrouter/meta/muse-spark-1.3"Dropping the openrouter/ prefix is the usual mistake here — a bare meta/muse-spark-1.3 reads meta as the provider, which is not the route you configured.
Self-Hosted: Hermes Agent
Hermes splits the provider out and takes the model name bare. It ships a dedicated provider for this family, meta-ai, which points at Meta's Model API and reads MODEL_API_KEY:
model:
provider: meta-ai
default: muse-spark-1.3Through OpenRouter instead, the provider becomes the aggregator and the model keeps the aggregator's slug:
model:
provider: openrouter
default: meta/muse-spark-1.3New sessions pick the change up on their own; hermes gateway restart only matters if you want to move sessions that are already open. One Hermes-specific quirk worth knowing on this family: the model rejects a reasoning effort of none, and Hermes maps a disabled setting to minimal rather than passing it through.
For credentials and the wider setup on each side, see OpenClaw + Muse Spark and Hermes + Muse Spark.
The Reasoning Setting Is the Whole Decision
Reasoning tokens are billed as output. A max reasoning mode is therefore a cost lever as much as a quality lever, and on an always-on agent the difference compounds across every message rather than showing up once.
- Leave reasoning low or default for the everyday driver: chat replies, lookups, routine tool calls, scheduled jobs. This is most of what an always-on agent does, and it is where 1.2-class behaviour was already sufficient.
- Turn it up for planning a multi-step job, debugging something that has already failed once, or any task where a wrong tool call has a side effect you would have to reverse.
- Watch the first week's spend. If you switch a busy agent to max reasoning and change nothing else, expect the bill to move. Compare against your previous week before deciding it was worth it.
Because pricing on this line has moved with each release, check the current rate on your provider rather than assuming 1.2's figures carried over. A model that is cheap per token but reasons for a long time is not automatically cheap per task.
When to Pick It
- Good fit: long-running agents that make many tool calls, work with large contexts, and where you care more about getting the step right than about latency.
- Poor fit: latency-sensitive chat where a reasoning pause is felt by the person waiting. Pair it with a fast Flash-tier model instead — see multi-model switching.
- Already on 1.2 and happy? There is no urgency. 1.3 is an addition to the line, not a deprecation of the previous one.
Frequently Asked Questions
Does Muse Spark 1.3 work on Hermes Agent?
Yes, and Hermes ships a dedicated provider for this family. Set model.provider to meta-ai — which points at Meta's Model API and reads MODEL_API_KEY — with model.default as the bare muse-spark-1.3. To go through the aggregator instead, use provider: openrouter with default: meta/muse-spark-1.3.
Is there a model called Muse 1.3?
No. The official name is Muse Spark 1.3; “Muse 1.3” is only shorthand. Muse Code is a separate product — Meta's terminal coding agent — not a model version.
Does max reasoning cost more?
Yes. Reasoning tokens bill as output, so it raises cost per task, and on an always-on agent that compounds across every message. Keep reasoning low for routine work and raise it for planning and debugging.