Meta Muse Models Hit Google Cloud Preview

// 18m agoMODEL RELEASE

Meta Muse Models Hit Google Cloud Preview

Meta’s Muse Wonder, Muse Spark 1.3, and Muse Ember are now available in public preview through Google Cloud’s global endpoint. All three support reasoning, function calling, structured output, and 1M-token contexts via an OpenAI-compatible Responses API.

// ANALYSIS

Google Cloud is turning Meta’s model lineup into a practical workload stack, from frontier coding to cheap agent execution.

–Muse Wonder targets complex software engineering and long-running agents
–Muse Spark 1.3 handles multi-step coding and agentic workflows at lower cost
–Muse Ember is optimized for high-volume subtasks, routing, extraction, and tool calls
–Google Cloud access reduces migration friction for teams already using Model Garden
–The shared 1M-token context and API surface make model routing easier to implement

// TAGS

muse-wondermuse-sparkmuse-emberllmagentcoding-agentcloudapi

DISCOVERED

18m ago

2026-11-09

PUBLISHED

1h ago

2026-11-09

RELEVANCE

9/ 10

AUTHOR

MetaforDevs

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

INFRA47m ago

Anomaly Teases Private Inference Stack

Anomaly’s OpenCode team previews a private project called Inference, described as “turning machines into tokens,” with a GPU-to-token workflow shown in its README. The post signals an internal inference infrastructure effort, but provides no public code, API, benchmarks, or launch details. [X post](https://x.com/thdxr/status/2119920964607643662)

BENCHMARK1h ago

Claude Fable 5.2 Takes Fifth on MacroscopeBench

Anthropic’s Claude Fable 5.2 scores 78.5 F1 at xhigh on MacroscopeBench, ranking fifth overall and narrowly trailing GPT-6.1 Sol. It also reportedly outperforms Claude Opus 5.5 at matched low, medium, and high effort levels.

INFRA1h ago

OpenAI Control Room adds 35-second kill switch

OpenAI’s Control Room centralizes authorization for model inference across research training, evaluation, and data-processing workloads, allowing people or monitoring systems to halt output within 35 seconds. It fails closed if the control service goes down and is rolling out across OpenAI’s research infrastructure.