kalomaze Calls Gemma 4 12B a Beautiful, Underappreciated Model With Fully Linear Multimodal Embeddings
kalomaze · x · 2026-09-28
kalomaze praised Gemma 4 12B as a "beautiful/underappreciated" model:
- Its multimodal embeddings are fully linear/standardized, with a clean architecture.
- It is reasonably deep, and at 12B the dense (vs MoE) choice still makes sense, unlike around 30B.
- His verdict: whatever you think of Google's RL, their pretrains are elite.
More from Models
- Just days after Anthropic's Opus 5.5 launch, 10 wild use cases emerge — FinanceYF5 · 2026-09-28
- User finds Opus 5.5 usage on Claude a far better deal than Astra on Codex — RexDouglass · 2026-09-28
- Dev builds private "gamefeel" benchmark to test LLMs on game design intuition — teortaxesTex · 2026-09-28
- Thesis chapter argues LLM logprobs can power probabilistic programming again — xuanalogue · 2026-09-28
- Blogger predicts all major US AGI labs will reach AGI in 2027, then ASI — Dr_Singularity · 2026-09-28
- Three underrated Gemini 3 Flash use cases: browser use, image gen, and grunt work — BuffaloConscious7919 · 2026-09-28