Meta releases lightweight AI model Muse Ember and launches API

AT A GLANCE

The report says Meta launches the lightweight model Muse Ember, focused on low-cost agent tasks, but its performance on knowledge and complex programming tests is weaker.

Article

Meta launches Muse Ember: 17 times cheaper than its own Spark, matching Claude Haiku 5.5

Dongcha Beating AI News Flash, Meta has released its lightest AI model, Muse Ember, specifically designed to handle tasks repeatedly performed by agents, such as tool calls, code modifications, and document analysis. Ember supports text, image, audio, and video inputs, with a context window of 1000000 Tokens.

On Artificial Analysis's Intelligence Index, Muse Ember and Claude Haiku 5.5 both scored 36 points, but the cost per task is only 0.14 USD, about one-third of the latter. Compared with Meta's own Muse Spark 1.3, Ember is only 1 point lower, yet its cost drops sharply to about 1/17.

The key to the low cost lies in Token utilization efficiency. The standard input and output unit prices of Ember and Haiku 5.5 are actually the same, but Ember's cache read price is 60% lower, and the output Tokens generated when completing evaluation tasks are only about one-third of Haiku's.

Meta plans to let the large model Muse Wonder handle planning complex tasks, and then hand a large amount of repetitive work to Ember. For example, let Wonder formulate a code migration plan, and multiple Ember agents modify files and run tests in parallel.

However, Ember also has obvious shortcomings. In Artificial Analysis's knowledge test, it answered only 21.1% of the questions correctly, and directly refused to answer another 64% of the questions. In the more complex engineering programming test FrontierCode 2.0, its score was 27.2%, lagging behind Haiku 5.5's 34.9%.

Ember is currently available on Muse Code and Meta Model API. The standard price is 0.10 USD per million input Tokens and 0.50 USD for output, but when the input of a single request exceeds 200000 Tokens, the unit price will increase to 3 times.

Original link https://m.theblockbeats.info/flash/376415

Key points

Meta has released the lightweight AI model Muse Ember for tasks such as tool calls, code modifications, and document analysis.
Muse Ember supports text, image, audio, and video inputs, with a context window of 1000000 Tokens.
On Artificial Analysis's Intelligence Index, Muse Ember and Claude Haiku 5.5 both score 36 points.
The report says Muse Ember's cost per task is 0.14 USD, about one-third of Claude Haiku 5.5's.
Muse Ember is 1 point lower than Muse Spark 1.3, with a cost of about 1/17 of it.
Muse Ember and Claude Haiku 5.5 have the same standard input and output unit prices, but the cache read price is 60% lower, and the evaluation output Tokens are about one-third of the latter's.
Meta plans for Muse Wonder to plan complex tasks, and then multiple Muse Ember agents to handle repetitive work.
In Artificial Analysis's knowledge test, Muse Ember answered 21.1% correctly, and another 64% of questions were refused.
In the FrontierCode 2.0 test, Muse Ember scored 27.2%, lower than Claude Haiku 5.5's 34.9%.
The report says Muse Ember is now available on Muse Code and Meta Model API.
The standard price is 0.10 USD per million input Tokens and 0.50 USD for output Tokens; after the input of a single request exceeds 200000 Tokens, the unit price increases to 3 times.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Muse Ember is a lightweight model for a large amount of repetitive agent work. The report emphasizes that it reduces task costs through fewer output Tokens and cheaper cache reads, rather than relying on lower standard input and output unit prices. It can handle multiple types of input, but its scores on knowledge Q&A and more complex engineering programming tests show that its capabilities have limitations.

Why it matters to readers

If the capabilities, pricing, and launch information in the report are accurate, developers can use Muse Ember to handle repetitive tool calls and file modifications, and leave complex planning to stronger models, thereby reducing the cost of some agent workflows.

Beginners can understand it as an AI model oriented toward low-cost execution work: suitable for repetitive tasks, but one should not judge whether it is suitable for knowledge Q&A and complex programming based only on price or a single composite score.

Risks and unknowns

The material does not include Meta's official announcement or product page, and the release, launch, pricing, and capability information still need to be confirmed by official sources.
The performance and cost comparison comes from evaluation results relayed in the report, without complete testing methods and conditions.
It does not explain Muse Ember's actual scope of availability, customer adoption, or real-world usage feedback.
The low accuracy on the knowledge test, the relatively high refusal rate, and the weaker complex programming score may limit its use scenarios.

Meta releases lightweight AI model Muse Ember and launches API

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center