Spark-X2.5-4B-ASHQ1-GGUF

Spark-X2.5-4B — ASHQ1 Quantization

ASHQ1 quantization of Spark-X2.5-4B (dense 4.1B, spark2_5 arch).

Status: published. Custom Spark-X2.5-4B imatrix (216 tensors, 36 layers). A second imatrix from the Qwen3.8 calibration set gives a byte-identical tier distribution at 4000 MiB — results are calibration-robust.

Architecture (why ASHQ1 fits as-is)

• Dense decoder, 36 blocks (per imatrix), hidden 4096, 32 heads GQA
• Hybrid attention: 1 full-attention + 3 sliding-window layers repeating — SWA affects KV-cache, not quantization; no special tensors
• Fused q_k_v_proj (like qwen35) → tied groups work out of the box
• MLP ffn_gate/ffn_up/ffn_down, RMSNorms (attn_norm per block)
• No MTP, no MoE, no SSM — plain dense path
• Vocab 131072, native 1M context, BF16, Apache-2.0

Imatrix tensor types (attn_gate, attn_qkv, attn_output, ffn_gate, ffn_up, ffn_down) are all covered by the existing class map — no new mappings needed. Preset spark2_5 added, detected via general.architecture metadata.

Quants

PPL measured with stock llama.cpp (llama-perplexity, -ngl 99 -c 1024 -b 512 --seed 7) on wikitext-2-raw wiki.test.raw. Absolute values are inflated (model reads raw text without its jinja template); relative comparison is valid — all files measured identically. Note: spark needs -c 1024 — at -c 512 the same Q8 scores ~2× worse (context fragmentation).

⚠️ PPL is a canary, not a rank column: PPL below the BF16 base flags over-confidence (sharpening), not quality. Cross-checking all PPL-only wins on a KLD-vs-BF16 battery (with Soulfate24) — rankings below are provisional until KLD-scored.

Notes

• Source BF16: XHToken/Spark-X2.5-4B-GGUF (BF16 conversion)
• Imatrix: custom Spark-X2.5-4B imatrix (216 entries, 36 layers)
• Runs on stock llama.cpp (native spark2_5 support since Sep 2026, no fork needed)
• Quantization and tuning by wepiqx