Skip to content

[cloud] MiniMax M3 fails to process tool result messages — stalls then returns empty response #16389

Open

Description

@rutis-agent

What is the issue?

When using MiniMax M3 via Ollama Cloud (minimax-m3:cloud) with OpenAI-compatible tool/function calling, the model fails to correctly process tool result messages. After receiving tool results, the model stalls for ~20 seconds and then generates an empty (or very short) text response instead of issuing further tool calls or providing a substantive answer.

DeepSeek models on the same pipeline (same client, same Ollama Cloud, same tool definitions) handle identical tool results correctly.

Environment

  • Model: minimax-m3:cloud
  • Client: Hermes Agent (CLI AI agent) routing through LiteLLM proxy -> Ollama Cloud
  • API format: OpenAI-compatible /v1/chat/completions with native function calling
  • OS: Ubuntu 24.04

Steps to reproduce

  1. Send a chat completion request with tools defined (e.g., a shell/terminal tool)
  2. The model correctly issues a tool call on the first turn
  3. Send the tool result back as a role: "tool" message with the corresponding tool_call_id
  4. Observe: model stalls for ~20s, then returns a short/empty response with no further tool calls

Relevant log output

From Hermes agent.log:

Tool: terminal completed (0.22s, 120 chars)
[~20 second gap]
Model responded with empty/no-tool-call text

For comparison, deepseek-v4-flash:cloud on the same pipeline:

Tool: terminal completed (0.11s, 848 chars)
[immediate response with continued tool usage]

Expected behavior

The model should process the tool result message and either:

  • Issue additional tool calls if needed
  • Generate a natural language response incorporating the tool results

Actual behavior

The model receives the tool result, stalls for ~20 seconds, then produces an empty or near-empty response with zero tool calls. It appears the model does not properly handle the role: "tool" message in the conversation history on Ollama Cloud.

Additional context

This issue does NOT appear to be a client-side problem — the same client, tool definitions, and message format work correctly with other models on Ollama Cloud (DeepSeek, Qwen, etc.). It seems specific to MiniMax M3 on Ollama Cloud.

Activity

  1. gotnochill815-web commented on Jun 1, 2026

    @gotnochill815-web

    I tested the same OpenAI-style tool-result flow locally on Ollama 0.24.0 in Google Colab.

    Local model tested:

    • qwen3:4b

    Result:

    • First assistant turn correctly emitted a tool call.
    • Sending the result back as a role: "tool" message with matching tool_call_id worked.
    • The model produced the final answer: The current time is 9:59 PM.

    So at least for local Ollama 0.24.0, the basic tool-result message loop works correctly. This supports the idea that the reported MiniMax issue may be specific to minimax-m3:cloud / Ollama Cloud rather than the general Ollama chat/tool message format.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions