Meta Muse Spark 1.4 surfaces in RocksDB pull request

An unreleased Meta model called muse-spark-1.4-internal has appeared in a public GitHub pull request on Meta’s RocksDB repository. First spotted by @LuminaBench, the reference points to the next version of Meta’s Muse Spark line already running in internal production tooling.

The pull request makes a small logging change to RocksDB’s experimental “live tuning” feature. With live tuning, an AI advisor watches a running database and recommends configuration changes. At Meta, that advisor is an LLM accessed through an internal service called muse. The change adds two fields to every recommendation: the name of the model that answered and its usage, including tokens and seconds. The stated goal is to compare a frontier model with small on-host models across the same tuning cycles, measuring both cost and outcome.

The model name appears in the PR’s test notes, which seem to have been generated automatically. They describe a companion internal change that was run against a real model, with the tuner service reporting muse-spark-1.4-internal for that run. A single call used 33,535 input tokens, including 17,780 cached tokens, and produced 8,169 output tokens, including 7,867 reasoning tokens. The call took about 65 seconds. That makes 1.4 a model already being called inside Meta’s infrastructure tooling rather than merely tested in isolation. The PR was closed without merging a few hours after it opened, which is common for Meta changes that land through its internal systems.

Meta released Muse Spark 1.3 on September 2, about a month after 1.2. It is available in Muse Code and the Meta Model API with a 1M-token context, priced at $1.25 per million input tokens and $4.25 per million output tokens. Meta pitched it as a big step up in coding and long-horizon agentic work, while using about 25% fewer tokens than 1.2. More than two months have passed since then, so a 1.4 model already in internal use suggests the next update may not be far off. Meta has not confirmed a name, timing, or feature set for Muse Spark 1.4.

Sources and related context

• RocksDB live tuning pull request: Supports: The PR describes the logging change and records the muse-spark-1.4-internal model ID and call usage in its test notes.
• Meta introduces Muse Spark 1.3: Related context: Meta's September 2 announcement explains the preceding release, its coding and agentic focus, availability, and lower token use than 1.2.
• Meta Model API Muse Spark pricing: Related context: Meta lists standard Muse Spark API pricing at $1.25 per million input tokens and $4.25 per million output tokens.