"Fast Mode" tier for Gemini 3.1 Pro

[eevmanu]

Is there any roadmap for a "Fast Mode" tier for the Gemini 3.1 Pro model (similar to Anthropic/OpenAI offerings)?

To be clear: I am not asking about routing requests to a smaller, faster model like flash or flash-lite.

I'm curious if there is any intention to offer an infrastructure-level "Fast Mode" for the flagship Gemini 3.1 Pro model itself (with high thinking level). The idea would be to offer a tier with significantly higher throughput/lower latency for the exact same Pro-level intelligence, but at a higher cost per token. This is similar to the fast/priority routing tiers currently offered by competitors like Anthropic and OpenAI.

I'm specifically wondering if recent algorithmic breakthroughs from Google Research might make this commercially viable soon. For example, the recent paper on TurboQuant demonstrates extreme memory compression for the Key-Value (KV) cache.

According to the research, TurboQuant enables quantization of the KV cache down to just 3 bits with zero accuracy loss, resulting in up to an 8x performance increase in computing attention logits on H100 GPUs.

Could we expect optimizations like TurboQuant, or other kernel-level/speculative decoding improvements, to be packaged into a premium "Fast Mode" for the Gemini 3.1 Pro endpoint in the near future?

Thanks.

[aniruddhaadak80]

Hey there. Thanks for sharing this.

This is a really interesting idea and something we have been discussing internally. We love seeing how the community is pushing the boundaries of the CLI.

We will definitely keep this in mind for future iterations.