Hi Gemma Team,
First off, thank you for the incredible work on the Gemma 4 series! The E2B and E4B models have been game-changers for light/mobile workloads.
However, there is a massive gap in the desktop consumer landscape: the 12GB VRAM bracket (e.g., RTX 4070 / 3060 12GB users), which represents a huge portion of local LLM enthusiasts and independent developers.
While standard 12B Dense models run well, they often trade off reasoning capability compared to 27B–35B models. On the other hand, running 27B+ models locally requires offloading or heavy quantization that severely degrades speed.
Proposal for Gemma 5:
Introducing an E12B (Effective 12B) architecture (e.g., a sparse/MoE or active-parameter architecture that yields 27B–35B intelligence while maintaining a 12B active VRAM/RAM footprint).
This would hit the absolute "sweet spot" for desktop hardware, allowing high-speed, high-context generation without forcing users into massive VRAM upgrades during current hardware/RAM market constraints.
Bringing the "Effective" architecture scaling up to the 12B class in Gemma 5 would empower millions of local deployment users.
Thanks for considering this feedback!
Hi Gemma Team,
First off, thank you for the incredible work on the Gemma 4 series! The E2B and E4B models have been game-changers for light/mobile workloads.
However, there is a massive gap in the desktop consumer landscape: the 12GB VRAM bracket (e.g., RTX 4070 / 3060 12GB users), which represents a huge portion of local LLM enthusiasts and independent developers.
While standard 12B Dense models run well, they often trade off reasoning capability compared to 27B–35B models. On the other hand, running 27B+ models locally requires offloading or heavy quantization that severely degrades speed.
Proposal for Gemma 5:
Introducing an E12B (Effective 12B) architecture (e.g., a sparse/MoE or active-parameter architecture that yields 27B–35B intelligence while maintaining a 12B active VRAM/RAM footprint).
This would hit the absolute "sweet spot" for desktop hardware, allowing high-speed, high-context generation without forcing users into massive VRAM upgrades during current hardware/RAM market constraints.
Bringing the "Effective" architecture scaling up to the 12B class in Gemma 5 would empower millions of local deployment users.
Thanks for considering this feedback!