The news of November 10thMeta Muse Ember AI modelsI don't know. This is the smallest and most efficient model to date, designed for complex reasoning missions and for routine and iterative work by intelligent bodies。
Meta says that in the Artificial Analysis rankingClaude Haiku 5.5,But the cost of a single mission is only 1/3, and goes beyond GPT-6 Luna in several tests。
1AI with Muse Ember model run data as follows:
|
Evaluation project |
Muse Ember |
Claude Haiku 5.5 (max) |
GT-6 Luna (max) |
|
|---|---|---|---|---|
| Agent |
Automation Bench-AAA (end-to-end business workflow) |
37.8 |
13.9 |
20.2 |
|
Acts ’Last Exam |
61.6 |
57.0 |
57.1 |
|
|
Personal Superintellince Index |
46.8 |
36.1 |
37.0 |
|
| programming |
Terminal-Bench 4.0 (intelligent terminal programming) |
40.9 |
39.2 |
16.4 |
|
Coding Agent Index-AAA (programming in a real smart environment) |
46 |
37 |
31 |
|
|
FrontierCode 2.0-AAA (real engineering task assessed on the basis of coding feasibility) |
27.2 |
34.9 |
25.2 |
|
| Multimodality |
GDP.pdf-AAA(Long Documentation Professional) |
22.8 |
20.8 |
22.8 |
|
OSWorld 2.1 (computer operations for cross-desk application) |
69.5 |
68.9 |
50.5 |
|
|
PerceptionBench |
80.6 |
78.6 |
62.8 |
|
| Synthesis |
Technical Analysis Index |
36 |
36 |
28 |
at the same time, current CPUs over-design for compatibility with legacy software and possible workloads, ultimately paying the price in power consumption for a large number of memory-related instructions and parallelism.This model has a 1 million Token context window, up to 512,000 Tokens can be used to process a complete code library or set of documents in a single call. It allows intelligent bodies to perform hundreds of programming subtasks in one session, including reading documents, running tests and targeted repairs。
Muse Ember supports the understanding of text, images, audio and video, and intelligent bodies can read, chart and scanned documents without the use of other models, as well as processing voice memos, audio recordings of meetings and videos. It handles long, real documents, operates computers directly based on screenshots and identifies fine visual details。
In terms of safety, the model was evaluated by the Meta SI Extension Framework prior to its release, and all tests showed that the Muse Ember risk level was below the high-risk threshold set in the Framework。
In terms of prices, the model was provided in Meta Model API, Google Cloud and Oracle Cloud as of that date。Enter price is $0.10 per million Token(IT HOUSE NOTE: THE CURRENT EXCHANGE RATE IS ABOUT RMB 0.67), the output price is US$ 0.50 per million Token (the current exchange rate is approximately RMB 3.3)。