Qwen3.8-27B-MTP-q27

Qwen3.8-27B-MTP -- q27 tiers (v2 recipes)

Quantized .q27 artifacts of Qwen/Qwen3.8-27B for the q27 engine. Converted from the official safetensors via BF16 GGUF; the trained MTP head is preserved (15 tensors at blk.64), so q27's speculative ladder works.

These recipes were derived fresh for this checkpoint by a per-tensor sensitivity sweep (q27 BUILDLOG 2026-08-14), not ported from the Qwen3.6 tiers. The 3.6 recipes measurably do not transfer: on 3.8 the q4-head error-cancellation family inverts, and the ssm_out Q8 promotion the 3.6 default/q6/q6k tiers carry is actively harmful at depth on this checkpoint (+14% dPPL at 24K-token single-pass vs +0.26% chunked -- it compounds through the GDN recurrence).

All four are smaller AND lower-PPL than the corresponding 3.6-recipe builds of this checkpoint. Long-context tail behavior verified per-position against a q6k reference (catastrophic counts 1-10 of 24,575 positions); needle and per-tier canonical digests below were derived on the exact files uploaded.

Recipes (repack.py): q4s-v2 = Q4 head + Q8 attn_output; v2.0 = Q8 attn_output; q6-v2 = + Q8 ffn_down; q6k-v2 = + Q8 ffn_up. Tokenizer is byte-identical to Qwen3.6's.

Serving notes: Qwen3.8 agentic use wants --think (thinking off, the model fails hidden tests it passes with thinking on) and q27 >= commit eb4a6b0, which auto-selects the model's trained XML tool dialect.

DFlash2 drafter packs (q27 >= v0.11.0)

q27's production decode path since 2026-09-08 is a DFlash2 block drafter (K=7). The drafter is z-lab/Qwen3.8-27B-DFlash2 (Apache-2.0); the two files below are that checkpoint packed into q27's weight format by tools/dflash2_pack.py, drafter weights only -- the engine supplies the embedding and the head from its own tensors, so the packs work with every tier above.

Serve with Q27_DFLASH2=<pack> Q27_BATCH=0 Q27_DFLASH2_RESERVE_GB=3 (single slot; the reserve keeps the KV pool off the drafter's VRAM) and the model card sampler; q27's tools/launch_q27_38.sh d2-pfx is the exact production recipe. Measured on Claude Code traffic (12 SWE-bench instances, 2026-09-09): 207 t/s aggregate decode, 3.89 tokens per round, 96.9% prefix reuse. To rebuild a pack yourself: python3 tools/dflash2_pack.py <drafter_dir> out.d2w [--q8] (torch + safetensors + numpy).

Gates (derived on the exact uploaded files)

Needle-in-haystack: 6/6 at ~120K tokens (q4s, default), ~100K (q6), ~40K (q6k), fp8 KV.

Per-tier greedy canonicals (--tokens "760,6511,314,9338,369" -n 128 --ctx 2048 --spec, RTX 5090/sm_120):

Note: canonicals gate change, not identity -- nearby tiers can share a trajectory on this prompt (q6-v2's digest matches the superseded v1 q6/q6k builds despite different tensors). File md5s are in CHECKSUMS.md5.