Competition between open-source AI and commercialization continues to intensify. Recently, Meta Superintelligence Labs officially released Muse Ember, the fourth and smallest high-efficiency multimodal reasoning model under the Muse brand to date. The model is purpose-built for repetitive subtasks that AI agents need to handle at high frequency in their daily operations.
Core Positioning: Focused on High-Frequency Agent Subtasks, Enabling Refined Division of Labor
In Meta's product matrix, each model plays a different role. Among them, the flagship Muse Wonder focuses on tackling the most complex planning and hard problems, Muse Spark handles everyday coding and agent workflows, while the new Muse Ember focuses precisely on large-scale repetitive coding and agent subtasks.
Officials explicitly noted that Muse Ember's core use cases mainly cover tool calls, coding subtasks, routing, extraction, and long document parsing. Through this refined division of labor, large models can support the massive daily operations of agents with greater efficiency.
Architecture Features: Native Multimodal Integration and a Million-Token Context Window
In terms of underlying architecture and multimodal capabilities, Muse Ember demonstrates a very strong overall processing level. The model natively supports multiple input modalities, including text, images, audio, and video. This means that when agents perform tasks, they can directly read screenshots, charts, scanned pages, voice memos, and even meeting recordings without additionally calling separate models, and can even operate desktop applications directly through screenshots.
In terms of processing capacity, Muse Ember has an ultra-large context window of up to 1 million tokens, and a single call can easily accommodate an entire codebase or a whole document collection. At the same time, its maximum output length can reach 512,000 tokens, up to 4 times that of similar competitors.
It should be noted that the model is fully hosted and runs only on Meta's managed infrastructure (namely Meta Model API and Muse Code). Its weights remain private, so there are no options for local GPU, laptop, or quantized deployment.
Benchmarks: Stunning Agent Task Performance, Slightly Behind in Code Merge Scenarios
In authoritative benchmark organizations and Meta's internal test suites, with maximum reasoning effort enabled, Muse Ember demonstrated excellent overall performance:
Agents and workflows (Agents): In the AutomationBench-AA test based on 657 Zapier business workflows, Ember achieved a task completion rate of 37.8%, far ahead of GPT-6Luna's 20.2% and Claude Haiku5.5's 13.9%. In the internal Agents' Last Exam test, Ember scored 61.6, also outperforming Haiku's 57.0 and Luna's 57.1.
Coding and development (Coding): In the Terminal-Bench4.0 test, Ember led with 40.9 points, ahead of Haiku's 39.2 and Luna's 16.4; it also ranked first with 46 points on the Artificial Analysis (AA) Coding Agent Index. However, in the FrontierCode2.0-AA test, which examines real-repository code merge capability, Haiku5.5 scored 34.9, outperforming Ember's 27.2.
Multimodality and perception: In the GDP.pdf-AA test, Ember tied with GPT-6Luna (22.8 points); on OSWorld2.1 (partial score), it narrowly led Haiku's 68.9 with 69.5 points; in the tool-integrated PerceptionBench test, Ember scored 80.6, higher than Haiku's 78.6 and Luna's 62.8.
In the overall Artificial Analysis Intelligence Index, Muse Ember scored 36, exactly on par with Claude Haiku5.5 (max effort) and ahead of GPT-6Luna's 28.
Cost-effectiveness: Same Base Pricing, Significantly Optimized Real-World Task Costs
In terms of public API pricing, Muse Ember's standard rates are consistent with competitors: $0.10 per million input tokens, $0.50 per million output tokens, and cached input at $0.004 per million tokens (lower than Haiku and Luna's $0.01).
Although the base unit price is similar, thanks to improved token consumption efficiency, the actual cost of executing a single task has been significantly optimized. Measured data shows that Muse Ember's cost per task under the highest reasoning setting is about $0.136, while Claude Haiku5.5's cost to complete comparable tasks is as high as $0.411, meaning its per-task cost has dropped to about one-third of its rival.
In addition, when processing ultra-long text, competitors often double prices or apply tiered surcharges when prompts exceed 10K or 272K tokens, while Meta has not set up a separate long-context surcharge tier for Ember, further strengthening its cost-effectiveness in long-text processing.
Ecosystem Access and Partners
Currently, developers can officially access and use Muse Ember through Meta Model API and Meta's coding agent tool Muse Code, with support for the standard Responses API. During the early testing phase, several well-known companies and technology platforms, including Box, Cline, Oracle, Cloudflare, and You.com, have already participated.
