How much ram is needed to run Qwen3 reasoning model?

The RAM requirements for running Qwen3 (and the recent Qwen3.5 update) depend significantly on the model size and the level of quantization (compression) used. Qwen3 introduces a "Thinking" mode for complex reasoning, which typically requires slightly more overhead than standard instruct models due to multistep processing.

Qwen3 & Qwen3.5 RAM Requirements

For local inference using 4-bit quantization (a common balance of speed and quality), the approximate memory needs are as follows:

Model Size	Min. RAM/VRAM (4-bit)	Recommended System RAM	Best Use Case
0.6B	~1 GB	8 GB	Mobile/Edge devices
1.7B	~1.5 - 2 GB	8 GB	Basic reasoning tasks
4B	~2.75 - 3 GB	10 GB	Logic and coding
8B	~5 GB	16 GB	General reasoning/chat
14B / 14.8B	~8 - 10 GB	24 - 32 GB	Advanced logic/science
30B-A3B (MoE)	~16.5 - 17.5 GB	32 GB+	Complex math & coding
32B	~20 - 22 GB	48 GB+	Heavy lifting/High accuracy
122B / 235B	~72 - 150 GB+	128 - 256 GB+	Enterprise-grade reasoning

Key Considerations for Reasoning

• "Thinking" Mode: When running Qwen3 in its native Reasoning (Thinking) mode, inference is slower because the model generates a chain of thought before providing an answer.
• Unified Memory: On MacBooks (Apple Silicon), the system RAM is shared with the GPU, making them ideal for larger models like the Qwen3-30B-A3B or Qwen3.5-122B if you have 32GB to 128GB of RAM.
• Context Length: High context (e.g., 128k or 1M tokens) drastically increases RAM usage. For example, a 7B model requires roughly 111 GB of memory to handle a full 1-million token context.
• Quantization: Using 8-bit (Q8) quants will roughly double the memory requirements listed above, while 1.5-bit or 2-bit quants can fit much larger models into less RAM at a significant cost to intelligence.

For most users wanting a solid reasoning experience, the Qwen3 8B on a 16GB RAM system or the 30B-A3B MoE on a 32GB RAM system are the "sweet spots" for performance and accuracy.

Finding Available Quants

Not every quantization is available for every model by default. To see what you can choose from:

• Go to the Ollama Library.
• Search for Qwen (e.g., qwen2.5 or qwen2.5-coder).
• Click the "Tags" tab. This lists every specific version and quantization level (e.g., 1.5b-instruct-q8_0, 7b-math-fp16) available to download.
• Use this format ollama run qwen2.5-coder:7b-instruct-q4_K_M