4 hours ago
36
Meta’s Muse Spark has quietly become one of the most-used AI coding models in the world. The model now ranks second on OpenCode Go’s weekly leaderboard, processing over 120 million sessions and 79 trillion developer tokens across the platform.
What Muse Spark actually does
Muse Spark emerged from Meta Superintelligence Labs as a purpose-built model for agentic coding. The model operates with a 1 million token context window, which means it can hold roughly the equivalent of several large codebases in its working memory at once.
The model series has evolved quickly since its April 2026 debut. Version 1.1 arrived in July, followed by version 1.2 on August 5, 2026, which was co-trained with Muse Code, Meta’s dedicated terminal agent. Version 1.3 landed in September, continuing the rapid iteration cycle.
OpenCode Go, the platform where Muse Spark has earned its number-two ranking, operates as a $10-per-month subscription service that gives developers access to various AI coding models optimized for terminal agents. The ranking is based on actual usage rather than synthetic benchmarks.
The numbers in context
The 120 million sessions metric suggests not just raw throughput but repeated, active engagement from developers who keep coming back to use the model. The research notes that independent verification of the claimed 120 million sessions and 79 trillion developer tokens remains elusive, though operational metrics from the platform confirm high token throughput across models.
Muse Spark has been performing strongly against competitors including DeepSeek, with Meta’s model sitting at the number-two spot on OpenCode Go’s weekly leaderboard.
Meta’s pricing for the model API sits at $1.25 per million input tokens and $4.25 per million output tokens at standard rates. The Contributor tier drops those prices to $0.10 and $0.20 per million tokens respectively, with developers on this tier contributing data that Meta uses for training purposes.
Meta’s bigger AI chess game
The Muse Spark series represents a notable strategic shift for Meta compared to its previous Llama model family. Where Llama was open-weight and broadly available, Muse Spark operates with closed weights, meaning Meta retains full control over the model’s architecture and parameters.
Meta has positioned the Muse Spark line under the banner of “personal superintelligence,” a term the company uses to describe AI systems capable of deeply personalized, highly autonomous reasoning.
OpenCode Go’s limited-region Muse Spark Contributor variants offer deeply discounted access in select regions, allowing Meta to grow its training dataset while building market share in additional geographies.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.
English (US) ·