Pick the region first. The model list is not the same everywhere.
Real-Time Voice Conversations
Real-Time Voice Conversations depend on where the build thinks you are. Global, Chinese mainland, and Russia do not share one fallback. Use Global if you are not on the mainland client and not in Russia.
gemini-2.5-flash is the primary Global model, on Google Vertex AI. As of v1.2.9 on 7 October 2026, that model is hitting usage-limit errors after demo players rose about twentyfold in a month. A manual tier increase was not available from Google Cloud. The demo is primarily on the OpenAI fallback while that lasts.
Global route
The server confirms that you own the game, returns a short-lived token, and voice does not pass through the Easy Fox server. Easy Fox does not opt in to OpenAI training. Realtime abuse logs can be kept up to 30 days.
Chinese mainland route
Mainland voice does not pass through the Easy Fox server either. ModelArk does not train on customer data. Safety-filtered input and output may be kept 180 days. Players on the Global build can still miss this route. Missing it leaves you on the Global chain, not on a hidden extra model.
Russia route
In patch 1.2.8, global AI services are blocked in Russia and a slower Russia-specific model is connected. In v1.2.9, alternative models for Russia are still being tested. The model’s name has not been announced.
Global token and what the server keeps
Ownership is checked, a short-lived token comes back, and the voice stream is not proxied through Easy Fox for the Global route. That is why a studio outage and a model outage are different failures. If the token step fails, you never reach gemini-2.5-flash or the Realtime fallback. If the token step works and the reply is empty, you are in the October model problem.
Training opt-in is off for OpenAI on this route. Abuse logs for Realtime can remain for 30 days. That is a retention window, not a transcript browser you can open from the pause menu. Do not expect the demo to hand you yesterday’s audio file.
The temporary model in patch 1.2.8 on 2 October 2026 was less capable and more reliable. v1.2.6 on 29 September had already moved the previous fallback up to primary because Gemini disruptions were frequent. Those two updates belong together: reliability was chosen over the preferred model before the twentyfold limit error on 7 October.
Mainland models and the Global-build miss
On the mainland route the order is Qwen3.5 Omni Flash, then Qwen3 Omni Flash, both through Alibaba Cloud Model Studio, then Doubao-Seed-2.0-mini on BytePlus ModelArk. The roles are primary service, first fallback, and second fallback. She tries them in that order.
Patch 1.2.8 gives the mainland QQ group 874276824 for players who downloaded the Global build and cannot reach the China model. Patch 1.2.4 on 26 September 2026 had already tried to route some of those Global-build players toward mainland model servers. If you are on the mainland and Global is the route answering you, that QQ group is the support path for the miss.
Retention on ModelArk is the 180-day window for safety-filtered input and output, and customer data is not used for training. Voice still does not pass through the Easy Fox server on this route. A long reply is still the v1.2.9 quality complaint. The route explains who hosts the model. It does not excuse a backwards turn.
Russia status without a fake model name
October gave Russia two statuses, and they are not the same status. On 2 October 2026 a slower Russia-specific model is connected because global AI services are blocked. On 7 October, alternatives for Russia are still being tested. Connected, and still testing, can both be true. The model’s name has not been announced.
Do not paste a Global model name into a Russia bug report and call it the Russia model. gemini-2.5-flash and the two Realtime Mini names are the Global chain. The Qwen and Doubao names are the mainland chain. Russia’s model name has not been announced. A slower answer is the symptom 1.2.8 already owns.
Reading about another region does not change your install or your network. Your install still uses the route for where the build thinks you are. If you are outside Russia and outside the mainland, judge the OpenAI fallback that is carrying the demo in v1.2.9.
Voice Quality and silence
Patch 1.1 on 9 June 2026 added a Voice Quality setting and overhauled her voice. Higher is more natural. Lower is faster. That setting is the lever when a reply is pretty and late. It is not the lever for a usage-limit error. A limit error is the host refusing the call. Quality is how the call sounds when the host accepts it.
Patch 1.2.3 on 8 September 2026 set TTS Quality to Low as part of the AMD crash work. If speech suddenly sounds flatter after that build, check the quality setting before you assume the model changed. The October fallback is a model change. The September Low flag is a performance flag. They stack. They are not aliases.
Earlier silence bugs are still useful as a checklist. Patch 1.0.8 on 4 April 2026 addressed occasional silence and the wrong spoken language. Patch 1.0.5 on 14 March 2026 addressed first-stage silence and language settings that did not apply. If she is mute on a current build, confirm you are on v1.2.9, confirm the microphone, then look at the route. A March silence fix does not explain an October token failure.