An internal Muse Spark 1.4 shows up in Meta's own RocksDB code, with a real model call logged, ten weeks after Spark 1.3

A merged commit in Meta's public RocksDB repository (meta/rocksdb) names a model called muse-spark-1.4-internal. The pull request, #15674, was opened at 01:52 UTC on November 13 and merged at 08:35 UTC. Its description says a companion internal change was run against a real model, and that the tuning service reported that model as muse-spark-1.4-internal. Meta has not announced Muse Spark 1.4. @LuminaBench spotted the pull request.

WHAT IS SOURCED

• The pull request and the merged commit (5d150fae), read on GitHub. The change adds two fields to RocksDB's experimental live-tuning feature so it records which model answered each tuning recommendation and what the call cost. The feature asks an LLM for those recommendations through what the pull request calls muse.
• The logged call, as quoted in the pull request: model muse-spark-1.4-internal, 33,535 input tokens (17,780 of them cached), 8,169 output tokens, 7,867 reasoning tokens, one model call, 64.96 seconds.
• The stated goal: comparing a frontier model with small on-host models on the same tuning cycles, by tokens, time and outcome.

WHAT IT FITS

• Meta shipped Muse Spark 1.1 on July 9, 1.2 on August 5 and 1.3 on September 2, about a month apart. No Spark point release has followed in the ten weeks since. In that time Meta shipped Muse Wonder (October 12), Muse Spark Open (October 15) and Muse Ember (November 9).
• The -internal suffix points to a build Meta is using in-house. ArtificialWatch's radar, which sweeps Meta's public model list, has not seen a Spark 1.4 id.

WHAT WE DO NOT KNOW

Whether muse-spark-1.4-internal ships publicly as Muse Spark 1.4, when, at what price, or how it compares with Spark 1.3. The pull request gives no benchmark. ArtificialWatch's Muse Spark 1.4 hunt fires the moment a 1.4 id appears on a list we sweep.