[Bug] Error from provider (Console Go): Upstream request failed — Kimi K3

[nyaa666]

Description

Kimi K3 model appears in the model selection list but throws "Error from provider (Console Go): Upstream request failed" when selected.

This error only happens with Kimi K3 — other models on Console Go work fine.

Key details:

• The model is listed and selectable, but fails on any request
• OpenCode Go limits are within normal range (~40% weekly, ~60% monthly) — not a quota issue
• Config files are clean: no custom top_p, temperature, or other parameter overrides in any config or agent files
• No top_p:0 found anywhere (unlike Error from provider (Console Go): Upstream request failed #37231 which was caused by that)
• Other Console Go models (e.g. deepseek-v4-flash) work without issues
• Changing API key does not help

OpenCode version

1.18.3

OS

Windows

Steps to reproduce

1. Open a new session
2. Select Kimi K3 from the model picker
3. Send any message
4. Receive: "Error from provider (Console Go): Upstream request failed"
   [...]
5. Switch to another model (e.g. deepseek-v4-flash) — works normally

Related

#37231, #37056, #36889

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• OpenCode Go: kimi-k3, kimi-k2.6, kimi-k2.7-code, glm-5.2, grok-4.5, qwen3.7-max, hy3-preview all fail with 'Upstream request failed' — non-standard 'mcp'/'system' fields rejected by strict provider validators #37771: Multiple Console Go models including Kimi K3 all fail with 'Upstream request failed' — covers the same symptom and appears to be the same root cause
• opencode-go (Console Go) provider returns 400/401/500 for subscribed models #37056: Console Go provider returning errors for subscribed models
• Error from provider (Console Go): Upstream request failed #37231: Same error message ('Error from provider (Console Go): Upstream request failed') — now closed

If your issue is already tracked in one of these, please add any additional context there rather than keeping a separate report.

[ctml91]

Any solutions for this yet?

[Caffeine19]

Same here in the VSCode Copilot.

      {
        "id": "kimi-k3",
        "name": "Kimi K3",
        "url": "https://opencode.ai/zen/go/v1/chat/completions",
        "toolCalling": true,
        "vision": true,
        "maxInputTokens": 1048576,
        "maxOutputTokens": 16384,
        "modelOptions": {
          "temperature": 0,
          "top_p": 0
        }
      },

[why6689]

The latest VS Code example may contain the key clue: the generated model descriptor still sends "top_p": 0, even though the original reporter did not set it in their own config.

That makes this look similar to #37231. I would first remove both top_p and temperature from the generated Kimi K3 request, start a fresh session, and retry. If that works, the fix likely belongs in the Console Go model
catalog rather than in each user's config.

For comparison, Kimi K3 is also available through an OpenAI-compatible endpoint on Routerra:
https://routerra.ai/register?aff=Uzwx