Gemini 3.5 Flash slow or malformed when generating large function-call arguments

[uriva]

Summary

gemini-3.5-flash can take 50-120s, return MAX_TOKENS, or return MALFORMED_FUNCTION_CALL when generating a large string inside a function-call argument.

This reproduces with a direct @google/genai call, without any agent framework or application wrapper.

Minimal repro gist:
https://gist.github.com/uriva/6bc9e20d1915c8d655ad3b3558796977

Environment

• SDK: @google/genai@2.4.0
• Runtime: Deno
• Model: gemini-3.5-flash
• API: Gemini API via API key
• Call: ai.models.generateContent(request)
• Streaming: not required to reproduce

Repro shape

The repro sends one user message and one declared tool:

• Tool: run_command
• Args schema:
  • command: string
  • params: any
• Prompt asks the model to call run_command with command: code_execution/write_file
• params.content should contain a complete single-file HTML landing page

Request size is very small:

• About 1.5KB serialized request
• About 268 prompt tokens in one observed run
• One user content item
• One function declaration
• No previous turns
• No thought signatures in input

Config:

thinkingConfig: {
  includeThoughts: false,
  thinkingLevel: ThinkingLevel.LOW,
},
maxOutputTokens: 16000,
toolConfig: { functionCallingConfig: {} },

Observed behavior

Across repeated runs of the same direct SDK request, outcomes vary:

1. Slow successful function call

   • ~50-60s latency
   • finishReason: STOP
   • returned run_command function call
   • generated ~48KB-56KB JSON function-call args

2. Token exhaustion

   • ~120s latency
   • finishReason: MAX_TOKENS
   • returned partial/large run_command function call args around ~56KB-60KB

3. Malformed function call

   • ~122s latency
   • finishReason: MALFORMED_FUNCTION_CALL
   • no content parts exposed by the SDK response

Example observed malformed run:

{
  "elapsedMs": 122344,
  "finishReason": "MALFORMED_FUNCTION_CALL",
  "usage": {
    "promptTokenCount": 268,
    "totalTokenCount": 268,
    "serviceTier": "standard"
  },
  "parts": []
}

Example observed slow successful run:

{
  "elapsedMs": 50557,
  "finishReason": "STOP",
  "usage": {
    "promptTokenCount": 234,
    "candidatesTokenCount": 12581,
    "totalTokenCount": 13288
  },
  "parts": [
    {
      "functionCall": "run_command",
      "argBytes": 48015
    }
  ]
}

Expected behavior

Large string arguments in function calls should either:

• complete reliably within reasonable latency, or
• fail quickly with an actionable error, or
• expose enough partial output/diagnostics to understand why the function call became malformed.

A large file write via a tool argument is a valid agentic workflow. The issue seems specific to generating large structured function-call JSON arguments, not to finding the tool.

Notes

• This is not caused by our app's agent wrapper; the gist calls GoogleGenAI directly.
• This is not caused by prior thought signatures; the minimal repro has no prior model turns.
• This is not about many tools; the minimal repro has one declared function.
• The model often does find the intended tool and starts generating params.content correctly, but generation is slow/flaky and sometimes ends as MAX_TOKENS or MALFORMED_FUNCTION_CALL.

[Varun-S10]

Hi @uriva, Thank for raising the issue and provide the reproducible code. We were able to reproduce. This is because asking the model to pass the complete HTML file inside a tool parameter makes it run very slowly and eventually fail with a malformed JSON error.
As a workaround, please update your prompt to output the HTML as regular text first instead of passing it as a tool parameter, and could you please try this and let us know if its works for you?

[uriva]

This is a coder agent, so how would that work? It needs to work within a harness that provides a file editing tool

[Varun-S10]

Hi @uriva, Thank you for bringing this to our attention. We truly appreciate you flagging this issue, we will file a bug internally.