Meta Muse Ember: Cheap Cloud Subagent or Local Model? (2026)

Muse Ember, released by Meta Superintelligence Labs on November 9, 2026, is a cloud-only model: no weights, so no Mac Studio distance can be calculated and nothing runs on your own hardware. It is built to be a cheap subagent. Artificial Analysis scores it 36 on its Intelligence Index at $0.14 per task, the same score as Claude Haiku 5.5 (max) at $0.41 per task (Artificial Analysis, 2026). That price changes the math for anyone who runs small models locally to save money on agent work.

What is Muse Ember?

Meta calls it its "smallest and most efficient model", the low end of a range that also includes Muse Spark and Muse Wonder (Meta, 2026). It is available as muse-ember-1.0 on the Meta Model API and in Muse Code, and nowhere else: Artificial Analysis lists availability as "Meta's first-party API and Muse Code". Meta has not published a parameter count.

The pricing, for prompts under 200k tokens of context, from Meta's model page:

Longer prompts cost more: above 200k tokens of context, Meta charges $0.30 per million input tokens, $0.012 cached and $1.50 output.

Can you run Muse Ember locally?

No. Muse Ember is cloud-only and Meta has released no weights, so there is no Mac Studio distance to quote: a 512GB Mac Studio has nothing to load, and no size is published to compare against. ModelFit therefore has no model page or hardware estimate for it. If you want a Meta model with downloadable weights, see our guide to Muse Spark Open on a Mac, which needs a 512GB Mac Studio at the very least.

How good is it for subagent work?

Strong on agent workflows, weak on knowledge. Artificial Analysis reports these points (Artificial Analysis, 2026):

• Muse Ember (max) is the cheapest model scoring 36 or higher, at $0.14 per task. Claude Haiku 5.5 (max) reaches the same score for $0.41.
• Its cached input is 60% cheaper than Claude Haiku 5.5 and GPT-6 Luna, and cached input is most of the cost in agent evaluations.
• In Muse Code it scores 46 on the Coding Agent Index, ahead of Claude Haiku 5.5 in Claude Code at 37.
• AA-Omniscience accuracy is 21.1%, the lowest of the similarly intelligent models compared, mostly because it declines to answer 64% of questions. That keeps its hallucination rate at 17.4%.

The pattern fits a subagent: a model that reads files, calls tools and edits code under a stronger planner, where world knowledge matters less than following steps. Give it retrieval or search when facts matter.

Cloud subagent or local model: which costs less?

It depends on volume and data. At $0.14 per Intelligence Index task, 1,000 such tasks cost about $140 (our arithmetic from the Artificial Analysis figure). Real agent tasks vary in length, so treat that as an order of magnitude. A local model has no per-token cost once you own the hardware, works offline, and keeps code and documents on your machine. That last point is often the deciding one for client work.

The trade-off is quality per dollar of hardware. These open models, each with a ModelFit page showing the memory it needs, are the usual local candidates for subagent duty:

The scores are not directly comparable: Muse Ember is not in the Epoch Capabilities Index, and those local models are not on the same Artificial Analysis chart. A practical setup uses both: a local model for private files and high-volume loops, Muse Ember for bursts when your machine is busy or too small. The ModelFit wizard shows which of these your Mac or GPU runs and at what estimated speed, and how much RAM you need for a local LLM explains the sizing.

FAQ

Is Muse Ember open source?

No. Meta offers Muse Ember only through the Meta Model API and Muse Code, under the ID muse-ember-1.0. No weights are published, so it cannot be downloaded or run on a Mac or GPU. Meta's open-weight releases, such as Muse Spark Open, are separate models.

How much does Muse Ember cost?

Under 200k tokens of context, $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens. Above 200k tokens the rates rise to $0.30, $0.012 and $1.50. Artificial Analysis puts its cost at $0.14 per Intelligence Index task.

Is Muse Ember better than Claude Haiku 5.5?

Artificial Analysis gives both 36 on its Intelligence Index at max effort, with Muse Ember at about a third of the cost per task. Muse Ember leads in Meta's own coding agent, while Claude Haiku 5.5 is stronger on knowledge, where Muse Ember scores 21.1% on AA-Omniscience.

What local model can replace Muse Ember as a subagent?

Qwen3.8 27B and Gemma 4 26B-A4B fit 24GB machines, Qwen3.6 35B-A3B fits 32GB, and Qwen3.5 9B fits 12GB GPUs. None is benchmarked on the same chart as Muse Ember, so test one on your own tasks. The ModelFit wizard ranks them for your hardware.

Sources

• Meta: Introducing Our Smallest, Most Efficient Model: Muse Ember
• Meta developer docs: Muse Ember
• Artificial Analysis: Meta's Muse Ember is the cheapest model scoring 36 on the Intelligence Index