共有:

Gemini 3.8 Live · Google DeepMind

Splitting the speed of thinking from the speed of talking.

Until now, one Gemini model handled both deep reasoning and casual back-and-forth. Google DeepMind has unveiled Gemini 3.8 Live, a model built specifically for conversation, alongside an Extended Thinking variant. Here's what that split means for voice assistant latency.

AI Navigate Editorial2026.09.166 min read

SINGLE MODEL (BEFORE) speaks one model does reasoning + chat reply (a beat late) GEMINI 3.8 LIVE speaks dedicated model instant reply
01
The Announcement

A model trained to be fast
at talking, not just thinking

Until now, one Gemini model carried both deep reasoning and real-time small talk.

Google DeepMind unveiled a new model built for spoken conversation, Gemini 3.8 Live, alongside an Extended Thinking variant, Gemini 3.8 Live Extended Thinking, according to its official site. Previously, Gemini didn't separate its reasoning-heavy model from the one answering everyday chat — the same underlying model did both. For voice-assistant use cases, that meant the longer the model thought, the longer the awkward silence before a reply.

Full specs and supported platforms for the Live lineup are documented on the Gemini model lineup page. The announcement reads as a shift within the Gemini brand itself — from "one model does everything" toward "pick a model built for the job," splitting by use case rather than by raw capability alone.

Gemini beforeGemini 3.8 Live
One model carries both reasoning and chatSplits into conversation-first and Extended Thinking tiers
Deeper thinking tends to delay the replyConversational speed is a design goal from the start
Voice assistant use often feels laggyOptimized for real-time back-and-forth

02
Why It Matters

Why a speed-dedicated model
was needed at all

Unlike text chat, voice carries its own constraint: the awkwardness of silence.

2 tiers
Live / Live Extended Thinking
Conversation-first
Stated as an explicit design goal
Same brand
A use-case split within the Gemini line

A text chatbot can get away with a bit of lag — users tolerate it. Spoken conversation can't: even a few hundred milliseconds of silence reads as "is it even listening?" Trying to guarantee that experience with a single general-purpose model means satisfying two competing demands at once — keep deep reasoning intact, while also making replies fast. Splitting into a dedicated model is a design call that resolves that tension by not asking one model to do both.

This mirrors a broader industry pattern: separating conversation-optimized and reasoning-optimized architectures signals that more companies are treating voice UI as a real product surface, not an afterthought. For developers who've mostly built text-first products, this adds a variable they haven't had to design for before — conversational latency as its own axis.

03
In Practice

Who this actually affects, and how

Voice product developers

For voice assistants, call-center copilots, and in-car voice UI where latency is the experience, switching to a conversation-dedicated model should show up directly as faster-feeling replies. Start by benchmarking response speed against your existing general-model implementation.

PMs and product designers

If you've been trying to get both "fast" and "deeply reasoned" out of one model, this opens room to rethink your roadmap around picking the right model per use case instead.

Long-form writers, mostly text

If your main use is long-form writing or code generation, this announcement barely touches your experience. Extended Thinking and the rest of the reasoning-first lineup remain the right tools for that work.


The awkward beat of silence in conversation
was never a problem text chat had to solve.


04
Risk & Outlook

What's still unproven

As of the announcement, it's still an open question how much reasoning quality is preserved once a model is split off to prioritize conversation. Design changes that favor speed typically trade off against depth of response. The existence of an Extended Thinking variant looks like a deliberate hedge against exactly that trade-off, but how well "fast without going shallow" actually holds up is something to reserve judgment on until real-world benchmarks are in.

In the near term, the practical next step for voice-first product teams is simply to A/B the response speed and quality against their current implementation. If the speed gain is genuinely noticeable, expect the "split into a dedicated conversation model" pattern to spread to other products that have been text-chat-first until now.