HF maintainer: EmbeddingGemma2 is actually a bidirectional continuation of Gemma 4
tomaarsen · x · 2026-10-29
Hugging Face transformers maintainer tomaarsen says EmbeddingGemma2 is not based on embeddinggemma-300m as docs suggest, but a bidirectional continuation of the Gemma 4 family. He points to modularembeddinggemma2.py in the repo, which generates the actual modeling file and shows heavy code reuse from Gemma 4 and Gemma 3. Questioner vncntrsnlt notes the causal-to-bidi adaptation is odd given Gemma 4 models aren't released yet and suggests inspecting weight shapes for lineage clues.
Related event: EmbeddingGemma2 Revealed as Bidirectional Gemma 4 Successor(3 posts)→
More from Models
- Artificial Analysis Intelligence Index v5: 60% now uses private test sets — ajratner · 2026-10-29
- Suhail launches Particle: particle-1 matches performance with ~45% fewer tokens — chrisalbon · 2026-10-29
- Musk praises Demis's Gemini 4 Carbon, SOTA on key benchmarks — elonmusk · 2026-10-29
- Exploring a whole repo with a 1M-context open model cost just $0.63 — HowDevelop · 2026-10-29
- ARC-AGI scores follow power law v2 = v1^4.53, explaining 98% of variance — typewriters · 2026-10-29
- Commentary: Google is now shipping frontier models every two weeks — heyshrutimishra · 2026-10-29