Changing the model

[juliendf]

Hi,

By default it is using gemini-2.5-pro but is there a way to use Flash instead ? it is much fatser.

thanks

[cevenogithub]

Hi, you need to set export GEMINI_MODEL="gemini-2.5-flash" to override the default 2.5 Pro model

[scidomino]

We use flash for some easy tasks intentionally. For example, we use flash to determine if the turn is over or if the model should keep talking. Such tasks would not benefit from using a more expensive model.

[varshneydevansh]

@scidomino This is what was my initial theory with CLI as I have seen this also with 2.5 pro in AIStudio. Where based on the complexity of the task model decides whether to use the Thinking or not. But yeah if I felt that it should have used thinking I explicitly mention it or frame my prompt a little different where it understood the need of thinking.

But, since in CLI it rarely uses 2.5 Pro and immediately switches to Flash I thought of looking it here.

[marty-sullivan]

@scidomino that makes a lot of sense, but we should definitely have a way to set that model explicitly too, i.e. gemini-cli should respect the GEMINI_FLASH_MODEL and GEMINI_EMBEDDING_MODEL to allow user to manage which models are being used for all background tasks.

[varshneydevansh]

In AiStudio it decides on the same model which means the model remains same but model is getting switched is entirely different.

Having /model command like we have a flag --model ?

Other commands we have -

│ > / │
╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
▲
help for help on gemini-cli
mcp list configured MCP servers and tools
memory manage memory. Usage: /memory <show|refresh|add> [text for add]
privacy display the privacy notice
quit exit the cli
stats check session stats. Usage: /stats [model|tools]
theme change the theme
tools list available Gemini CLI tools
(15/16)

[x7dl8p]

Hi, you need to set export GEMINI_MODEL="gemini-2.5-flash" to override the default 2.5 Pro model

what would be the model name for gemeni-2.0-flash ?, this name showing not found.

[andreaskmueller]

So I am having the opposite issue. It starts with Pro but then automatically reverts to flash (2.5).

My gemini.md under my username (C:\User\USERNAME\gemini.md

has

---

model: gemini-2.5-pro

And it will start in it then revert to flash...

[cisterciansis]

i have the same issue as you. I have to use an API key. I set up my .env file to have export GEMINI_MODEL="gemini-2.5-pro" and when i launch gemini it loads in with pro but then quickly reverts to flash. I have billing set up on this account and api key. I also pay for google one AI. Im not sure why it keeps reverting back to flash. even using gemini --model gemini-2.5-pro first loads in pro but then switched to flash. its very odd and frustrating considering I pay monthly to use it and also went through the steps to set up a paid API key and set up my gemini env file etc.

[cisterciansis]

lol okay i fixed it. i had to refresh my ai studio api page and then link a new billing account to it that i set up. after that i reloaded gemini cli and it seems to stay on pro now. phew ...

[varshneydevansh]

Same for me it starts with pro 2.5 and then switches to Flash 2.5
gemini-2.5-flash (90% context )

[marty-sullivan]

it seems to use both models in some capacity, but I can't seem to find a way to override this. You can override the main model with GEMINI_MODEL environment variable. Looking at this file, you can see it also has a default for "flash" and "embedding" models:

gemini-cli/packages/core/src/config/models.ts

Lines 7 to 9 in edd69cb

	export const DEFAULT_GEMINI_MODEL = 'gemini-2.5-pro';
	export const DEFAULT_GEMINI_FLASH_MODEL = 'gemini-2.5-flash';
	export const DEFAULT_GEMINI_EMBEDDING_MODEL = 'gemini-embedding-001';

But it doesn't seem like setting GEMINI_FLASH_MODEL or DEFAULT_GEMINI_FLASH_MODEL does anything, it still uses gemini-2.5-flash for some things, even with gemini-2.5-pro set to the main model. I can see the calls to both pro and flash in my logs.

[oooseun]

you might have authenticated under the api key where it default to flash if you did not setup any $$$

[andreaskmueller]

I figured it out. Check the api keys and pick the one that says tier 1 (paid) Thank you Andreas [cid:logo-small_d7a6505a-c44e-4afc-9bf7-457dc874e18a.png] Andreas Mueller Chief Operating Officer Downstreem LLC 617.895.6615<tel:617.895.6615> | ***@***.******@***.***>. On Jun 28, 2025, at 07:21, Douglas ***@***.***> wrote:  [EXTERNAL EMAIL] DO NOT CLICK links or attachments unless you recognize the sender and know the content is safe. lol okay i fixed it. i had to refresh my ai studio api page and then link a new billing account to it that i set up. after that i reloaded gemini cli and it seems to stay on pro now. phew ... — Reply to this email directly, view it on GitHub<#1824 (reply in thread)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BDVUCP5MQ7BOKOTZQ2W4ZGD3F2QHVAVCNFSM6AAAAACAFNK24SVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTGNRQGUZDKMI>. You are receiving this because you commented.Message ID: ***@***.***>

[marfago]

I think it should be possible to switch the model from the console. For example "/model". You may want to switch from one to the other just in different sessions.

[varshneydevansh]

I might explore this as this seems to be at-least do not introduce any sort of circular dependency or major code refactor.

[DavidMetcalfe]

Agreed. This would be much more user friendly than using environmental variables.

[marfago]

It's not just more user friendly. It is a different use case. Having a command to switch model in the CLI will allow user to switch within the same conversation which I think would not be possible with an env variable.

[eelang-sc]

In fact it would be best if there was a hotkey for this.

[MarcinOrlowski]

hotkey is irrelevant as you currently cannot talk to model unless it's your turn. This is pretty annoying if you know how handy this is (and Claude Code does that). It also supports /model which is extremely handy + you can somehow control model switching from your prompt with certain wording (so you can i.e. by default run on faster and cheaper model but once you use think hard or so, it should (and in most cases does so) use thinking model for that prompt handling then switch back.

Please team - take a closer look at Claude Code features. They do it right way.

[itse4elhaam]

Is there a way to automate this?

If a model reaches its limit with 429 automatically switch to the next model?

[acoliver]

So I had sent a pull request for this exact thing like on the first day, and they rejected it. So I created a multi-model multi-provider fork:
https://github.com/acoliver/llxprt-code.
npm install -g @vybestack/llxprt-code

It still works for Gemini, but we did some changes that apply even to Gemini users:

• /model - you can switch to Flash
• No nextSpeakerChecker - right now, during a chat turn, it sends multiple requests to Flash to ask if Gemini Pro should keep speaking. It seems not to be needed and it sends the entire chat every time (burning a lot of tokens).
• You can customize the prompts (the prompts are REALLY long)

Luckily, due to the magic of LLMs, we are able to keep up with Gemini-CLIs release (0.1.17 - we added GPT-OSS support). We have a cherrypicker script and the LLM fixes the conflicts!

So if you want to use these things like /model, custom prompts, changing the model settings (temp etc) -- you can have it if you swim downstream :-)

I actually think Google is making the right decision. They have a particular vision and a very Gemini-centric one. For people who are doing more agentic stuff or need to control the budget -- a community project downstream can satisfy them while the Gemini Code Assist team focuses on their overall roadmap.

[Explorer-Dong]

use --model param before start gemini-cli, for example:

gemini --model gemini-2.5-flash

[x7dl8p]

model name for gemeni 2.0 flash ?, this config shows not found, but if i change it to 2.5 it works.

{
  "ideMode": true,
  "selectedAuthType": "oauth-personal",
  "hasSeenIdeIntegrationNudge": true,
  "model": "gemini-2.0-flash"
}

[aymuos15]

Having the option to change this once starting would be nice

[x7dl8p]

there is.
in the .gemini folder under settings.json, add model, something like this :

{
  "ideMode": true,
  "selectedAuthType": "oauth-personal",
  "hasSeenIdeIntegrationNudge": true,
  "model": "gemini-2.5-flash"
}

[aymuos15]

Is this is the equivalent of /model of claude code?

[x7dl8p]

model name for gemeni 2.0 flash ?, this config shows not found, but if i change it to 2.5 it works.
this is in .gemeni/settings.json directory.

{
  "ideMode": true,
  "selectedAuthType": "oauth-personal",
  "hasSeenIdeIntegrationNudge": true,
  "model": "gemini-2.0-flash"
}