Meta has released Muse Ember 1.0, which it describes as its smallest and most efficient model. Press and model trackers date the launch to 9 November 2026. Meta’s own developer post, “Start building with Muse Ember, our smallest, most efficient model”, is undated on the page, so that date rests on reporting rather than on Meta’s page. For research offices, the practical points are that access is through a paid, hosted service, and that Meta’s page does not say the weights are released.
What Meta says it released
According to Meta’s developer post, Muse Ember is built for high-volume execution: coding subtasks, tool calls, and document summarisation and extraction. It accepts text, image, audio and video inputs. It is available in all Muse Code subscription plans and through the Meta Model API, which exposes an OpenAI-compatible endpoint under the model ID muse-ember-1.0.
Cost
Meta lists API pricing per million tokens as follows.
| Input size | Input | Output |
|---|---|---|
| Up to 200k input tokens | $0.10 | $0.50 |
| Above 200k input tokens | $0.30 | $1.50 |
Meta describes the pricing as “optimized for low price-per-task”. Prices can change, so check Meta’s current rate card before budgeting a project.
Openness and licence
Meta’s developer post does not state whether model weights are released or under what licence. One trade report says the weights remain private, but that is a secondary account. Until Meta publishes licence terms, treat Muse Ember as a hosted, closed-access model. That differs from the open-weight approach associated with Meta’s earlier Llama releases. A hosted model also means the terms of the Meta Model API and Muse Code plans, not an open licence, govern academic use.
Benchmark claims (reported by Meta)
The following figures are Meta’s own, drawn from its developer post, and we have not independently reproduced them.
- Artificial Analysis Intelligence Index score of 36.
- AutomationBench-AA: 37.8, against 13.9 for Claude Haiku 5.5 and 20.2 for GPT-6 Luna (both at max reasoning).
- Terminal-Bench 4.0: 40.9, against 39.2 for Claude Haiku 5.5 and 16.4 for GPT-6 Luna.
- Coding Agent Index-AA: 46, against 37 and 31.
- FrontierCode 2.0-AA: 27.2, below Claude Haiku 5.5 at 34.9; GPT-6 Luna scored 25.2.
Meta calls the model “competitive in its class”, and the results above show it does not lead every benchmark.
What it means for researchers and research offices
- Cost: low per-token prices make bulk tasks such as extraction from large document sets cheaper to run, though costs depend on volume and any long-context surcharge.
- Reproducibility: a hosted model without released weights can change or be retired, so record the exact model ID (
muse-ember-1.0), access date and settings in methods sections. See our guidance on reproducibility. - Data governance: sending research or participant data to a hosted API needs review under institutional data-handling and ethics rules before use.
- Disclosure: use of generative AI in research work should be disclosed according to journal and funder policies.
