More accurate temperature support for Gemini models (e.g. Gemini 3 Pro needs to support temperature adjustment)

[Slava-256]

What's happening?

I generated a script to test temperature support for gemini models:

import os
import requests

api_key = os.environ["GEMINI_API_KEY"]
base_url = "https://generativelanguage.googleapis.com/v1beta"

models_response = requests.get(f"{base_url}/models?key={api_key}")
models = models_response.json()

# Scan for max model name length
max_name_len = max(len(model["name"]) for model in models.get("models", []))

for model in models.get("models", []):
    model_name = model["name"]
    padded_name = model_name.ljust(max_name_len)
    
    has_temperature = "temperature" in model
    has_max_temp = "maxTemperature" in model
    
    if has_temperature:
        default = model["temperature"]
        max_temp = model.get("maxTemperature", "unspecified")
        output = f"{padded_name}: temperature supported (default={default}, max={max_temp})"
        
        # Test models where max temperature is unspecified
        if not has_max_temp:
            supported_methods = model.get("supportedGenerationMethods", [])
            if "generateContent" in supported_methods:
                payload = {
                    "contents": [{"role": "user", "parts": [{"text": "hi"}]}],
                    "generationConfig": {"temperature": 2.0}
                }
                r = requests.post(
                    f"{base_url}/{model_name}:generateContent?key={api_key}",
                    json=payload
                )
                if r.status_code == 200:
                    output += " - test request with temperature=2.0 succeeded"
                else:
                    output += " - test request with temperature=2.0 failed"
            else:
                output += " - skipped test (generateContent not supported)"
        
        print(output)
    else:
        print(f"{padded_name}: temperature not supported")

And I got the following output:

models/gemini-2.5-flash                             : temperature supported (default=1, max=2)
models/gemini-2.5-pro                               : temperature supported (default=1, max=2)
models/gemini-2.0-flash                             : temperature supported (default=1, max=2)
models/gemini-2.0-flash-001                         : temperature supported (default=1, max=2)
models/gemini-2.0-flash-exp-image-generation        : temperature supported (default=1, max=2)
models/gemini-2.0-flash-lite-001                    : temperature supported (default=1, max=2)
models/gemini-2.0-flash-lite                        : temperature supported (default=1, max=2)
models/gemini-exp-1206                              : temperature supported (default=1, max=2)
models/gemini-2.5-flash-preview-tts                 : temperature supported (default=1, max=2)
models/gemini-2.5-pro-preview-tts                   : temperature supported (default=1, max=2)
models/gemma-3-1b-it                                : temperature supported (default=1, max=unspecified) - test request with temperature=2.0 succeeded
models/gemma-3-4b-it                                : temperature supported (default=1, max=unspecified) - test request with temperature=2.0 succeeded
models/gemma-3-12b-it                               : temperature supported (default=1, max=unspecified) - test request with temperature=2.0 succeeded
models/gemma-3-27b-it                               : temperature supported (default=1, max=unspecified) - test request with temperature=2.0 succeeded
models/gemma-3n-e4b-it                              : temperature supported (default=1, max=unspecified) - test request with temperature=2.0 succeeded
models/gemma-3n-e2b-it                              : temperature supported (default=1, max=unspecified) - test request with temperature=2.0 succeeded
models/gemini-flash-latest                          : temperature supported (default=1, max=2)
models/gemini-flash-lite-latest                     : temperature supported (default=1, max=2)
models/gemini-pro-latest                            : temperature supported (default=1, max=2)
models/gemini-2.5-flash-lite                        : temperature supported (default=1, max=2)
models/gemini-2.5-flash-image                       : temperature supported (default=1, max=1)
models/gemini-2.5-flash-preview-09-2025             : temperature supported (default=1, max=2)
models/gemini-2.5-flash-lite-preview-09-2025        : temperature supported (default=1, max=2)
models/gemini-3-pro-preview                         : temperature supported (default=1, max=2)
models/gemini-3-flash-preview                       : temperature supported (default=1, max=2)
models/gemini-3-pro-image-preview                   : temperature supported (default=1, max=1)
models/nano-banana-pro-preview                      : temperature supported (default=1, max=1)
models/gemini-robotics-er-1.5-preview               : temperature supported (default=1, max=2)
models/gemini-2.5-computer-use-preview-10-2025      : temperature supported (default=1, max=2)
models/deep-research-pro-preview-12-2025            : temperature supported (default=1, max=2)
models/gemini-embedding-001                         : temperature not supported
models/aqa                                          : temperature supported (default=0.2, max=unspecified) - skipped test (generateContent not supported)
models/imagen-4.0-generate-preview-06-06            : temperature not supported
models/imagen-4.0-ultra-generate-preview-06-06      : temperature not supported
models/imagen-4.0-generate-001                      : temperature not supported
models/imagen-4.0-ultra-generate-001                : temperature not supported
models/imagen-4.0-fast-generate-001                 : temperature not supported
models/veo-2.0-generate-001                         : temperature not supported
models/veo-3.0-generate-001                         : temperature not supported
models/veo-3.0-fast-generate-001                    : temperature not supported
models/veo-3.1-generate-preview                     : temperature not supported
models/veo-3.1-fast-generate-preview                : temperature not supported
models/gemini-2.5-flash-native-audio-latest         : temperature supported (default=1, max=2)
models/gemini-2.5-flash-native-audio-preview-09-2025: temperature supported (default=1, max=2)
models/gemini-2.5-flash-native-audio-preview-12-2025: temperature supported (default=1, max=2)

I see some discrepancies with what Big-AGI exposes in model option dialogs. For example, temperature adjustment is greyed out for Gemini 3 Pro Preview.

Is it possible/not too difficult for Big-AGI to pull temperature support for Gemini models from what it gets from "/models" API endpoint?

If a given model has "temperature" defined, then the model supports temperature, and "temperature" value is the default. If "maxTemperature" is defined, then that would specify the max value for the temperature, otherwise assume it is 2.0.

If "temperature" is not defined for a given model, then it can be assumed that temperature is not supported and the slider can be greyed out.

Where does this happen?

Big-AGI Pro (big-agi.com)

Impact on your workflow

Medium - Workaround exists

Environment (if applicable)

No response

Additional context

No response

[github-actions]

Thanks for the detailed investigation with the API testing script!

Assessment

You're correct - the Gemini API does return temperature and maxTemperature fields for each model, and Big-AGI currently doesn't use this dynamic information. Instead, Big-AGI uses a hardcoded LLM_IF_HOTFIX_NoTemperature flag for Gemini 3 models.

Current Behavior

In src/modules/llms/server/gemini/gemini.models.ts:145, Gemini 3 models are configured with:

const IF_30 = [LLM_IF_HOTFIX_NoTemperature /* vital: the Gemini 3 Developers Guide strongly recommending keeping it at 1 (aka not setting it) */, ...IF_25];

This was based on Google's Gemini 3 Developer Guide which recommends:

"For Gemini 3, we strongly recommend keeping the temperature parameter at its default value of 1.0. Changing the temperature (setting it below 1.0) may lead to unexpected behavior, such as looping or degraded performance, particularly in complex mathematical or reasoning tasks."

The Issue

While Google recommends keeping temperature at 1.0, they don't disable it. Users should be able to adjust it if they understand the implications. Your testing confirms that temperature up to 2.0 works fine for Gemini 3 models.

Proposed Solution

There are two approaches:

Option A: Remove the hotfix flag (quick fix)
Simply remove LLM_IF_HOTFIX_NoTemperature from IF_30 in gemini.models.ts. This would immediately enable temperature adjustment for Gemini 3 models with the existing 0-2 range.

Option B: Dynamic temperature support from API (more comprehensive)
Use the temperature and maxTemperature fields from the Gemini /models API response to:

1. Determine if temperature is supported (presence of temperature field)
2. Set the max temperature range dynamically based on maxTemperature (or default to 2.0 if unspecified)

This would require changes to:

• gemini.models.ts: Update geminiModelToModelDescription() to pass through temperature metadata
• llm.server.types.ts: Add maxOutputTemperature field to ModelDescription_schema
• LLMParametersEditor.tsx: Use the dynamic max value for the temperature slider

Recommendation

I suggest Option A as an immediate fix since:

1. All Gemini models that support temperature have maxTemperature: 2 according to your testing
2. The UI already supports temperature up to 2.0 via the "overheat" button
3. Option B can be implemented later as a broader enhancement

Would you like me to implement Option A?

[Slava-256]

FWIW, I actually want to experiment with increasing the temperature, not decreasing. For example, with Gemini 2.5 Pro I am getting better results with my prompts when I set temperature to values higher than 1.0.

I will go out on a limb and speculate that Big-AGI's motivation to opt for disabling the temperature for Gemini 3.0 Pro was based on general default for temperature in Big-AGI being 0.5. So someone experimenting with gentle temperature adjustments for Gemni 3.0 Pro would've ended up with temperatures being significantly lower than 1.0, potentially causing problems google's developer guide is warning about.

But if you pull the default temperature values from the "/models" API endpoint response for Gemini models, then the default for Gemni 3.0 Pro would be 1.0, and a user doing gentle adjustments would have much lower risk running into issues.

[enricoros]

@Slava-256 limb correct -- ok for me to:

1. change the baseline temperature to 1.0 for Gemini
2. relax the fix
3. push changes
4. work on dynamic from the API, if possible
5. Leverage the script above (in progress)

[enricoros]

Part 1 Request injection (global, non per-model)
I've added to the inspector the box at the bottom where one can insert { "json": 42 } values.

For instance in the screenshot below I set the Gemini temperature to 3, and all the models have rejected it.

This may be useful to you for when you want to set values globally and all the models will honor them as this is the last operation being done before sending out the request.

[enricoros]

@Slava-256 Gemini models have the full 0...2 temperature range now, and default to 1.

After testing the models I found your same conclusions:

• most Gemini LLMs have temp=1, max=2
• the Image LLMs (Nano Banana / Pro) have temp=1, max=1, but it can be forced to max=2
• Gemma models do not declare the temperature, but support temp=1, max=2
• and non-llm, such as imagen and veo, do not have the temperature

The current default is to have temperature=1, and the range [0 ... 2] for all models.

The temperature can be changed both on the Model Properties Dialog, or the Chat Right Pane. If not shown in the chat right pane, the temperature can be enabled like this:

[Slava-256]

@enricoros Thank you for the quick change! And thank you for further testing! I had a hunch you cannot always trust what the API returns. I am curious to experiment with Nano Banana Pro with temps > 1.0 :-)

[enricoros]

Gemini Models - parameter-sweep report

=== SWEEP SUMMARY (2026-02-05) ===

gemini (gemini)
  models/gemini-3-pro-preview (Gemini 3 Pro Preview)
    temperature                ✅ [0..2]
    gemini-thinking-level      ✅ [low, high] | ❌ [minimal, medium]
    gemini-thinking-budget     ✅ [1024, 16384, 24576, 32768, 65535] | ❌ [0]
  models/gemini-3-pro-image-preview (Nano Banana Pro)
    temperature                ✅ [0..2]
    gemini-thinking-budget     ✅ [0, 1024, 16384, 24576, 32768, 65535]
  models/gemini-3-flash-preview (Gemini 3 Flash Preview)
    temperature                ✅ [0..2]
    gemini-thinking-level      ✅ [minimal, low, medium, high]
    gemini-thinking-budget     ✅ [0, 1024, 16384, 24576, 32768, 65535]
  models/gemini-2.5-pro (Gemini 2.5 Pro)
    temperature                ✅ [0..2]
    gemini-thinking-budget     ✅ [1024, 16384, 24576, 32768] | ❌ [0, 65535]
  models/gemini-2.5-pro-preview-tts (Gemini 2.5 Pro Preview TTS)
    temperature                ✅ [0..2]
  models/gemini-2.5-flash (Gemini 2.5 Flash)
    temperature                ✅ [0..2]
    gemini-thinking-budget     ✅ [0, 1024, 16384, 24576] | ❌ [32768, 65535]
  models/gemini-2.5-flash-image (Nano Banana)
    temperature                ✅ [0..2]
  models/gemini-2.5-flash-preview-09-2025 (Gemini 2.5 Flash Preview 09-2025)
    temperature                ✅ [0..2]
    gemini-thinking-budget     ✅ [0, 1024, 16384, 24576] | ❌ [32768, 65535]
  models/gemini-2.5-flash-preview-tts (Gemini 2.5 Flash Preview TTS)
    temperature                ✅ [0, 0.5, 1.5, 2] | ❌ [1]
  models/gemini-2.5-flash-lite-preview-09-2025 (Gemini 2.5 Flash-Lite Preview 09-2025)
    temperature                ✅ [0..2]
    gemini-thinking-budget     ✅ [0, 1024, 16384, 24576] | ❌ [32768, 65535]
  models/gemini-2.5-flash-lite (Gemini 2.5 Flash-Lite)
    temperature                ✅ [0..2]
    gemini-thinking-budget     ✅ [0, 1024, 16384, 24576] | ❌ [32768, 65535]
  models/gemini-2.0-flash-exp-image-generation (Gemini 2.0 Flash (Image Generation) Experimental)
    temperature                ✅ [0..2]
  models/gemini-2.0-flash-lite-001 (Gemini 2.0 Flash-Lite 001)
    temperature                ✅ [0..2]
    gemini-thinking-budget     ✅ [0] | ❌ [1024, 16384, 24576, 32768, 65535]
  models/gemini-2.0-flash-001 (Gemini 2.0 Flash 001)
    temperature                ✅ [0..2]
    gemini-thinking-budget     ✅ [0] | ❌ [1024, 16384, 24576, 32768, 65535]
  models/gemma-3n-e4b-it (Gemma 3n E4B)
    temperature                ✅ [0..2]
  models/gemma-3n-e2b-it (Gemma 3n E2B)
    temperature                ✅ [0..2]
  models/gemma-3-4b-it (Gemma 3 4B)
    temperature                ✅ [0..2]
  models/gemma-3-27b-it (Gemma 3 27B)
    temperature                ✅ [0..2]
  models/gemma-3-1b-it (Gemma 3 1B)
    temperature                ✅ [0..2]
  models/gemma-3-12b-it (Gemma 3 12B)
    temperature                ✅ [0..2]