Back to guide/General Productivity

Open-Source Model Optimization Prompt

DeepSeek's open-source nature means many users run it locally. This prompt works well because DeepSeek understands its own architecture. Asking for copy-paste config values instead of general advice produces actionable output for the specific hardware setup.

deepseek-promptsmodelVarianthardwarequantization

Edit View

Prompt

I'm running DeepSeek {{modelVariant}} locally with the following setup:

**Infrastructure:**
- Hardware: {{hardware}}
- Quantization: {{quantization}}
- Context length: {{contextLength}}
- Inference framework: {{inferenceFramework}}

**Problem:** {{performanceProblem}}

**Current configuration:**
{{currentConfig}}

Analyze my setup and recommend optimizations:
1. **Memory optimization** — How to reduce VRAM usage without significant quality loss
2. **Speed optimization** — Inference latency improvements
3. **Quality optimization** — Sampling parameters for my use case: {{useCase}}
4. **Configuration changes** — Specific settings to adjust with before/after expected impact

Provide exact config values I can copy-paste, not general advice.

Variables to customize

{{modelVariant}}{{hardware}}{{quantization}}{{contextLength}}{{inferenceFramework}}{{performanceProblem}}{{currentConfig}}{{useCase}}

Why this prompt works

Save this prompt to your library

Organize, version, and access your best prompts across ChatGPT, Claude, and Cursor.

Related prompts

Task Decomposition Agent

Forcing the agent to plan before acting prevents premature execution and wasted steps. Explicit dependency mapping enables parallel execution and catches logical gaps early.

Tool Selection Agent

The ReAct pattern (Reason + Act) creates an explicit reasoning trace that improves tool selection accuracy. The error-handling rule prevents infinite retry loops.

Prompt Compressor

Explicitly requiring all functional requirements to be preserved prevents the model from over-compressing and losing critical instructions.

Memory Management Agent

Explicit memory read/write instructions create agents that improve over time. Categorization keeps memories organized, and the deduplication rule prevents context bloat.