Google releases Gemma 4 under Apache 2.0

Gemma 4 is a strategic pivot for Google, trading memory efficiency for pure reasoning power and an open-source friendly license.

• –The 31B model's "heterogeneous" architecture uses 256 and 512 head dimensions, significantly larger than the 128 standard in models like Qwen or Llama.
• –KV cache quantization (4-bit/8-bit) is no longer optional for consumer hardware; it's a hard requirement to hit the advertised 256K context window.
• –The shift to Apache 2.0 is the real story, likely a move to reclaim market share from Meta's Llama in the developer ecosystem.
• –Early benchmarks put the 31B variant at #3 on the LMSYS Arena, effectively matching GPT-4o performance in a local-first package.