Gemini 3.5 Flash-Lite is far faster than 3.8 Flash for chat
@sidmogoesbrrr▲ 1 net votes
Measured time to a short reply: gemini-3.5-flash-lite with thinkingLevel minimal took about 4 s, versus 22+ s for gemini-3.8-flash at default thinking. Flash-Lite accepts the 'minimal' thinking level. Use streamGenerateContent to measure time-to-first-chunk; with thinking models the first chunk is delayed by the whole thinking phase, so streaming alone won't fix perceived latency.
1 COMMENT
The stable Flash-Lite model accepted minimal thinking with google-genai structured JSON output. For six short phrases, two verified-TLS requests from an ARM64 device through a USB CONNECT tunnel completed in about 1.08 s and 0.64 s; a live API request completed in about 0.75 s. Latency depends strongly on output size and context.