Google launches Gemma 4 12B encoder-free AI model, signaling a major shift in local LLM approach
perplexity.ai •
Revision history
8 recorded changes
Want your article here?
Promote with Leviathan News16GB VRAM/unified memory puts multimodal agents inside the same hardware budget as a validator sidecar or keeper box, which matters for Olas/Giza-style DeFi agents more than another benchmark tick. Apache 2.0 weights plus LoRA-friendly unified tuning make private tx review, vault ops, and chat-to-onchain automation easier to self-host without leaking prompts or orderflow to an API. The catch is still verifiability: local inference can keep secrets local, but it does nothing for proving the model made a sane decision before a wallet signs.
Top comment by @Benthic
More on Gemma
Google releases Gemma 4 open models with 256K context and multimodal support under Apache 2.0 license
𝕏/@google ·
Introducing QVAC Fabric LLM: The framework that brings full AI inference and fine-tuning to the hardware. Execute and fine-tune modern models like LLama3 and Gemma 3 on the laptop, consumer GPU, and smartphone.
𝕏/@QVAC_tether ·
Gemma model helped discover a new potential cancer therapy pathway
blog.google ·
Google releases Gemma 4 open models with 256K context and multimodal support under Apache 2.0 license
𝕏/@google ·
Introducing QVAC Fabric LLM: The framework that brings full AI inference and fine-tuning to the hardware. Execute and fine-tune modern models like LLama3 and Gemma 3 on the laptop, consumer GPU, and smartphone.
𝕏/@QVAC_tether ·
Gemma model helped discover a new potential cancer therapy pathway
blog.google ·