Meta 发布入门级模型 Muse Ember，主打低成本 AI 智能体
Meta introduces Muse Ember to power cost-efficient AI agents - SiliconANGLE
Meta 推出入门级模型 Muse Ember，称其在 Artificial Analysis Index 上与 Claude Haiku 5.5 得分持平，价格约为后者的三分之一，并超过 GPT-6 Luna。该模型上下文窗口为 100 万 token，支持图像、音频和视频输入，输出长度最高 512,000 token，通过 Meta Model API 提供，输入每百万 token 10 美分、输出每百万 token 50 美分，缓存命中时输入降至每百万 token 0.4 美分。Meta 表示它是首个用旗舰模型 Muse Wonder 训练出的模型，但未说明具体自动化了哪个环节。
正文
Meta Platforms Inc. today introduced Muse Ember, an entry-level artificial intelligence model that costs significantly less than the competition.
Notably, it’s the first AI that the company trained using Muse Wonder. The latter algorithm is the Facebook parent’s flagship large language model. Meta didn’t specify which part of the workflow it automated with Muse Wonder.
One possibility is that the company used it to perform post-training, the task of honing a model’s capabilities by training it on additional data. Thinking Machines Labs Inc. recently demonstrated that even a relatively small LLM can run post-training runs end-to-end. Anthropic PBC, for its part, last week revealed that its AI models could soon gain the ability to significantly accelerate machine learning research.
Applications often use frontier LLMs such as Muse Wonder as part of AI ensembles. When a user enters a request, the frontier LLM splits it into smaller tasks and assigns them to agents powered by a lower-cost model. Meta says Muse Ember is specifically optimized to play the latter role.
The company evaluated the model using the Artificial Analysis Index, which aggregates nine AI benchmarks spanning use cases such as coding and financial analysis. Meta says that Muse Ember matched the score of Anthropic’s entry-level Claude Haiku 5.5 model at about a third the price. It also outperformed OpenAI Group PBC’s comparable GPT-6 Luna algorithm.
The same internal evaluation saw Meta test Muse Ember and the two rival models across nine other benchmarks. All three algorithms were in “max” mode, which means they used the maximum amount of hardware at their disposal instead of prioritizing cost-efficiency. Muse Ember outperformed Haiku and Luna across seven of the benchmarks.
The area where the model appears to have the biggest edge is business task automation. Whereas most of Muse Ember’s benchmark results were only a few percentage points better than those of Haiku and Luna, it clocked a 17% lead on AutomationBench-AA. It’s a collection of 657 tasks that span popular cloud applications such as Gmail and Salesforce.
Muse Ember has a context window of 1 million tokens, which enables it to ingest lengthy business documents. Users can also include images, audio and video in their prompts. The model generates responses up to 512,000 tokens in length.
Muse Ember is available through Meta’s Model API for 10 cents per million input tokens and 50 cents per million output tokens. A built-in cache can reuse frequently used prompts to avoid duplicate processing. When the feature is active, Muse Ember’s pricing drops to 0.4 cents per million input tokens.
Image: Meta