Quantized weights15.98 GB+ 0.64 GB format overhead
strongcodingagentic-codingtool-uselong-context
Qwen3-Coder 30B-A3B Instruct
Strong coding-oriented choice based on model-specific evidence; benchmark scores are shown individually rather than merged.
Model profile
Parameters30.5B3.3B active
Maximum context262,144
LicenseApache-2.0
Released2025-07-31
32K / one-agent memory trace
Using GGUF Q4_K_M. This is fit math, not a fixed “billions to GB” lookup.
KV cache3 GBarchitecture-aware
Total with workspace + safety23.42 GB48 GB recommended RAM
KV cache uses layers × KV heads × head dimension × K/V × cache precision × context × concurrency.
Architecture metadata
| Layers | 48 |
|---|---|
| Attention / KV heads | 32 / 4 |
| Head dimension | 128 |
Evidence records
inference
Measured on a Qwen3 30B-A3B Q4_K_M checkpoint, not the exact Coder checkpoint; used only for a wide interpolation.