---

license: apache-2.0

language:

• en

base_model: XHToken/Spark-X2.5-1.7B

pipeline_tag: text-generation

library_name: gguf

tags:

• spark
• sliding-window-attention
• hybrid-attention
• 1m-context
• reasoning
• coding
• agent
• quantized
• quantization
• gguf
• llama-cpp
• imatrix
• hybrid-quantization
• ASHQ1
• priority-queue

---

Spark-X2.5-1.7B — ASHQ1 Quantization

> 🧪 New experimental quant available: BF-RELIEF-1000.gguf — beats all regular quants at 1000 MiB (PPL 66.70). More coming soon.

ASHQ1 quantization of Spark-X2.5-1.7B (dense ~2B, spark2_5 arch) — the little sibling of Spark-X2.5-4B, same hybrid attention (1 full + 3 sliding-window layers), native 1M context.

> Status: published. Dedicated 1.7B imatrix (168 tensors, 28 layers) from Qwen3.8 calibration set. F16 reference PPL on calibration set: 12.21.

Architecture (why ASHQ1 fits as-is)

• Dense decoder, spark2_5 arch string — same spark2_5 ASHQ1 preset as the 4B (layer count auto-detected: 28 blocks here vs 36 on 4B)
• Fused q_k_v_proj, MLP ffn_gate/ffn_up/ffn_down, RMSNorms
• No MTP, no MoE, no SSM — plain dense path
• Native 1M context, BF16, Apache-2.0

Quants

| File | Size | PPL (wiki.test.raw, ctx 1024) | Description |

|:-----|:----:|:-----------------------------:|:------------|

| BF-RELIEF-1000.gguf | 1006 MiB | 66.70 ± 0.68 | 🧪 Experimental — best at 1000 MiB, details coming |

| Spark-X2.5-1.7B-BF16-Q8_0.gguf | 1.7 GB | 60.91 ± 0.62 | Uniform Q8_0 reference |

| Spark-X2.5-1.7B-BF16-ASHQ1-1000.gguf | 1006 MiB | 71.50 ± 0.74 | Recommended: ASHQ1 bottom-up, output/token_embd pinned at Q5_K |

| Spark-X2.5-1.7B-BF16-ASHQ1-1000-FREE.gguf | 1006 MiB | 73.14 ± 0.76 | Experimental: --free-pins (pins in open budget fight) — loses, pins validated |

| Spark-X2.5-1.7B-BF16-ASHQ1-1000-EMBD8.gguf | 1006 MiB | 83.42 ± 0.86 | Experimental: embd pinned at Q8_0 — starves the rest, worst of the three |

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.

> ⚠️ 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. See the TOX lesson on our Ornith repo: a −0.145 PPL "win" with KLD 0.2894 (degraded).

Notes

• Source BF16: XHToken/Spark-X2.5-1.7B-GGUF (BF16 conversion)
• Runs on stock llama.cpp (native spark2_5 support since Sep 2026, no fork needed)
• Quantization and tuning by wepiqx