gemma4:26b (MoE) returns completely empty response (no content, no reasoning) and stops early on long system prompts

What is the issue?

When using the Gemma 4 MoE model (gemma4:26b) via the native /api/chat endpoint, the model returns a completely empty response (empty content, missing reasoning, and done_reason: "stop") if the system prompt exceeds roughly 500 characters.

This is not the known issue with the OpenAI /v1/chat/completions endpoint where the output is moved to the reasoning field (as reported in #15288). In this case, the model evaluates ~49 tokens and stops immediately, producing absolutely no output.

If the system prompt is short (e.g., < 200 chars), the model works correctly and produces output. The same long prompt works perfectly on Dense models like gemma4:31b and gemma-d-cc:latest.

Steps to reproduce

Create a long system prompt (e.g., 2000+ characters). You can use any filler text or a complex system prompt.

# Create a 2000-character dummy system prompt
SYS_PROMPT=$(printf "System instruction. %.0s" {1..100})

# Send request to native API
curl -s http://localhost:11434/api/chat -d '{
  "model": "gemma4:26b",
  "messages": [
    {"role": "system", "content": "'"$SYS_PROMPT"'"},
    {"role": "user", "content": "Who won the 2025 NCAA mens basketball championship?"}
  ],
  "stream": false
}' | jq .

Result:

{
  "model": "gemma4:26b",
  "created_at": "2026-04-08T19:59:20.718104Z",
  "message": {
    "role": "assistant",
    "content": ""
  },
  "done": true,
  "done_reason": "stop",
  "total_duration": 1265782584,
  "load_duration": 137602292,
  "prompt_eval_count": 1423,
  "prompt_eval_duration": 192706834,
  "eval_count": 49,
  "eval_duration": 911844667
}

Notice eval_count is 49 and content is "".

What I have ruled out

• Modelfile/Parameters: I tested identical blobs with varied Modelfiles (removing RENDERER, PARSER, tweaking num_ctx, temperature, top_k, etc.). The bug persists across all variants.
• Ollama Version: Reproduced on 0.20.0 and 0.20.3.
• Prompt Content: The bug triggers even with purely all-ASCII junk (Lorem Ipsum) if it reaches the length threshold.
• Dense Models: gemma4:31b and gemma4:e4b handle the exact same prompt correctly. This bug is isolated to the MoE architecture or its specific quantization in the library.

Environment

• Ollama version: 0.20.3
• Model: gemma4:26b (SHA: 7121486771cbfe218851513210c40b35dbdee93ab1ef43fe36283c883980f0df)
• OS: macOS 15 (Apple Silicon Mac Studio)
• Configuration: flash_attention: false, num_parallel: 1