Architecture: Qwen3 MoE (Mixture of Experts)
Total Parameters: ~30B
Active Parameters: ~3.3B (A3B)
Max Context Length: 256K (Note: Requires significant RAM/VRAM at full context; lowering to 8K-32K is recommended for standard hardware).

Available Quantizations

File Name	Bit Depth	Description
Qwen3-30B-A3B-Instruct-2507-REAM-heretic-Q4_K_M.gguf	4-bit	⭐ Recommended. Best balance of speed, RAM usage, and intelligence.
Qwen3-30B-A3B-Instruct-2507-REAM-heretic-Q5_K_M.gguf	5-bit	Higher quality, slightly larger RAM requirement.
Qwen3-30B-A3B-Instruct-2507-REAM-heretic-Q6_K.gguf	6-bit	Very high quality, near unquantized performance.
Qwen3-30B-A3B-Instruct-2507-REAM-heretic-Q8_0.gguf	8-bit	Extremely high quality, large file size.
Qwen3-30B-A3B-Instruct-2507-REAM-heretic-Q3_K_M.gguf	3-bit	Smallest file size, noticeable intelligence loss. Use only if heavily RAM constrained.

Prompt Format (ChatML)

This model uses the ChatML format. If you are using a UI like LM Studio or Ollama, it should detect this automatically from the GGUF metadata.

text

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Hello!<|im_end|>
<|im_start|>assistant