| # Qwen3 tiny fixture configuration decision | |
| The local 0.6B GGUF identifies itself as `general.architecture = qwen3` and | |
| contains 28 uniform decoder layers. Its characteristic geometry is query width | |
| twice the hidden width, KV width equal to hidden width, Q/K per-head norms, and an | |
| FFN ratio of three. | |
| The tiny fixture preserves those relationships with hidden width 128, four query | |
| heads, two KV heads, head dimension 64, FFN width 384, two layers, context 128, | |
| and RoPE base 1000000. All matrix quantization axes are multiples of 32. | |
| The fixed Hugging Face implementation is Transformers 5.14.1 | |
| `Qwen3ForCausalLM`. The trainable parameter count is 508,800. The generated | |
| package uses BF16 and tied token/output embeddings. The GGUF is derived from the | |
| same BF16 values through a temporary F32 serialization and pinned Q4_0 | |
| quantization; the F32 intermediate is not retained. | |