Flash Onyx 2.1: One Day Later – Model Spent 400 Tokens Thinking and Returned an Empty String

Description and use

Flash Onyx 2.1, after compressing the system prompt and disabling the thinking channel, became faster, stopped returning empty responses, and produced more accurate, concise answers.

Problem

The Flash Onyx 2.1 model, running Gemma 4 12B via Ollama, was slow and sometimes returned empty responses due to a large, inefficient system prompt and misconfigured thinking channel.

Outcome

After compressing the system prompt and adjusting the thinking channel, Flash Onyx 2.1 achieved faster response times, eliminated empty responses, and produced concise yet complete answers on most test prompts.

Workflow

• 1. Measure prompt evaluation time and token usage via API timing fields.
• 2. Identify excessive prompt length and internal reasoning causing delays.
• 3. Compress the system prompt by removing redundant rules and shortening sentences.
• 4. Re‑run tests on a set of seven representative prompts to evaluate performance.
• 5. Adjust brevity rules to preserve substance while keeping answers concise.
• 6. Test position of rules; move critical ones to the end of the prompt to trigger changes.
• 7. Verify that the model no longer returns empty strings and that response quality improves.
• 8. Deploy the updated model via Ollama command line.
• 9. Monitor performance and iterate if necessary.

Evidence

prompt_eval_count 27000ish prompt_eval_duration ~290s eval_count 40

Four hundred tokens generated. Zero characters out.

From 903 lines to 516, 27,000 tokens to 14,200, prefill from about 290 seconds to about 152.

Final tally: 2.1 won six, tied one, lost none, and its answers came to 70% of Onyx 2's length.