arpicato/qwen38-27b-8plus9-splice

Qwen3.8 Splice Package (canonical: IQ4_XS-gate/up + F16 alpha/beta)

Package files:

spliced-iq-f16ab.gguf       # canonical build (sensitivity-guided: F16 ssm_alpha/beta)
splice-iq-f16ab-metadata.json  # provenance
spliced-iq.gguf             # previous canonical (Q8_0 alpha/beta)
splice-iq-metadata.json

On HF these are published as spliced-iq.gguf + splice-iq-metadata.json (no generic spliced.gguf name anywhere — it is not descriptive). The baseline (unsloth/Qwen3.8-27B-GGUF @ 4ca7207...) is NOT ours and is never stored here; the A/B driver fetches it from upstream (see bench/ab/run_ab_pod.sh).

Student tensor scheme (matches Unsloth per-tensor mix): Q4K attngate/attnqkv/ffndown, IQ4XS ffngate/ffnup (their imatrix), Q5K ssmout, Q80 ssm_alpha/beta, F32 norms/bias/conv. Official llama.cpp Qwen3 V-head reorder applied to student linear-attention tensors.

Layer order:

source 0..7, student(8+9), source 10..64 shifted to output 9..63

History (all removed 2026-09-04):

• —spliced.gguf (F16 student, +342 MB over baseline) — void comparison.
• —spliced-q4k.gguf (Q4_K-only student) — valid A/B 290/548 vs baseline 315/548, recorded in results/ab-20260903T195909Z/ARTIFACT.md, HF commit ed9fc71bad7719806ec252a8eaafe667673688fe.
• —baseline.gguf — byte-identical Unsloth copy, not a result.

Verify uploads:

sha256sum -c SHA256SUMS

The spliced files retain raw GGUF tensor payloads for untouched source tensors. Student matmuls are quantized to match the baseline scheme (Q4K modal; ssmout Q5K; ssmalpha/beta Q80; norms/bias/conv F32). The -iq variant uses IQ4XS (with unsloth's imatrix) for ffngate/ffnup, mirroring the baseline's per-tensor mix. Official llama.cpp Qwen3 V-head reorder is applied to student linear-attention tensors.

Layer order:

source 0..7, student(8+9), source 10..64 shifted to output 9..63

• —spliced.gguf was removed 2026-09-04 (old F16-student build, superseded)
• —baseline.gguf was removed 2026-09-04 — it is the Unsloth source artifact, not ours; fetch it from unsloth/Qwen3.8-27B-GGUF revision 4ca720788d1e01f1bff70c033e0d0028fd02e502. The A/B driver does this automatically (see bench/ab/run_ab_pod.sh).

Verify uploads:

sha256sum -c SHA256SUMS