Qwen3.5-4B-DFlash GGUF

Q8_0 GGUF conversion of z-lab/Qwen3.5-4B-DFlash for use with Magnitude. This is a DFlash draft model for speculative decoding with Qwen/Qwen3.5-4B; the tokenizer metadata comes from that target.

Converted with llama.cpp at revision 43fe9c64281ef735046adc025e9e7559a1f659a5.

Artifact

• File: Qwen3.5-4B-DFlash-Q8_0.gguf
• Quantization: Q8_0
• SHA-256: 8864ab4b1d147278c3cf0806ec9aa9f0db644987f084856317e3186bf433280a
• Source revision: 9a1996ccf887b79ab3af4fcbf8c1d1f4b5658bcf

Draft metadata

• dflash.attention.causal = [true, true, true, true, true, false]: the first five layers use causal attention; the last uses bidirectional attention.
• dflash.sample_from_anchor = false: proposals come from mask rows, starting after the anchor.
• Derived from the checkpoint configuration and Z-Lab’s reference implementation; verified against reference outputs.