📊 BENCHMARKS & EVALS
- Community PCA head lifts 577M EmbeddingGemma-2 to #1 on MTEB Audio leaderboard
A community modder pushed Google's 577M EmbeddingGemma-2 from rank #26 to #1 on the MTEB Audio (MAEB) leaderboard by adding a 2.4 MB frozen PCA head with zero training.- Aimed at local/edge developers running audio and text-to-audio embedding search on laptops and edge devices.
- Weights released as BrunoHays/embeddinggemma-2-nullspace-audio; MAEB score rose from 49.3% to 54.4% (+5.1 points), moving from #26 to #1 on audio-only and #16 to #1 on full MAEB, beating 8B-9B multimodal models.
- The entire addition is a 2.37 MB safetensors file of frozen PCA matrices — no weights were trained.
- Base model is 577M parameters with 768-dimension vectors, roughly 15x smaller than 9B Omni models, so it runs on consumer hardware.
- Shows training-free post-hoc heads can make a small edge-sized embedder outperform much larger audio models.