Qwopus3.6-27B-v2-MTP-q27

Qwopus3.6-27B-v2-MTP -- q27 format

Repack of Qwopus3.6-27B-v2-MTP (a Qwen3.6-27B-MTP fine-tune by Jackrong) for the q27 inference engine. Same recipe family as the vanilla base repack; the fine-tune's higher MTP acceptance makes it ~35% faster under q27's speculative decoding on identical hardware (5.8 vs 3.6 tokens/round on real agent-transcript replay).

Files

One tokenizer serves every tier; it is byte-identical to the base model's.

Quality tiers

Measured 2026-07-27 on an RTX 5090. wikitext-2 PPL via --nll wiki.test.qwopus.i32 --nll-chunk 2048 --ctx 2048 (297,054 tokens, 145 chunks). auto-ctx is the window --ctx auto resolves to at fp8 KV on a 32 GB card.

* q8's PPL is measured on the --nll-serial path, because the batched path OOMs at 28.45 GB of weights on a 32 GB card. Serial computes logits through the decode kernel rather than prefill, so it was validated first: the default tier reads 7.1154 serial vs 7.1132 batched, an offset of +0.0022 (0.031%). q8's batched-equivalent is therefore ~7.0541. q8 also fails to boot as a server on 32 GB (it was tried at --ctx 8192), so on a 5090-class card it is a perplexity reference only.

Two things worth knowing before you pick a tier

1. The --q4-head recipe family that wins on vanilla LOSES here. On the vanilla repack, the q4s / q5f / q6f tiers exploit an error-cancellation effect between a single Q4 lm_head and Q4 residual writers -- q4s there measures BETTER than the larger default. That does not reproduce on this fine-tune. Size-matched pairs, where the only difference is the Q4 head:

q5f and q6f were built and measured but are deliberately not published here: on this checkpoint each is beaten by a smaller tier already in this repo. If you are porting tier guidance from the vanilla repo, do not -- the recipe family is checkpoint-specific, and every fine-tune needs its own sweep.

2. Tiers buy perplexity margin, not benchmark wins. HumanEval+ is 30/30 on every tier that boots. LCB v6 spans only 19-21/30, and that spread is noise, not ranking: of the 18 problems that failed at least once, only 2 failed on all six tiers, while 16 failed on some and passed on others. The best-PPL tier (q6k) fails problems the worst-PPL tier (q4s) passes. Pick by VRAM and context, not by the PPL column.

Bits are also non-monotonic in PPL: q8 (8.10 bpw) is beaten by q6k (6.80) and effectively tied by q6 (6.00). The same non-monotonicity appears on vanilla.

Usage

git clone https://github.com/signalnine/q27 && cd q27 && make
# bare defaults are the measured Claude-Code stack (fp8 KV, PMIN 0.5,
# MAXD auto7, fast-head, no-think); --ctx auto-sizes to VRAM
./build/q27-server qwopus-27b-mtp.q27 qwopus-27b-mtp.tok --port 8081

Verify a download:

md5sum -c CHECKSUMS.md5 --ignore-missing

Determinism

Canonical gate (--tokens "760,6511,314,9338,369" --ctx 2048 --spec -n 128, md5 of the generated: line, RTX 5090): 4c4120c72056aba2bc2d2561471eafce -- bitwise across the full spec-depth ladder (d4/d5/d6/d7/auto). That anchor is for the default tier; each tier is its own bitwise artifact and has its own canonical.

Performance (RTX 5090, 2026-07-09)

• Real-CC-transcript replay @25.8K: 211.9 t/s fixed depth-5, 220.7 t/s adaptive ladder (+4.2%).
• Live Claude Code serving: ~220 t/s decode bursts, 95% prompt prefix-cache.
• Prefill: identical to base (same weights shape) -- 128K in ~59.4s.

Provenance

Qwen3.6-27B-MTP by Alibaba/Qwen (Apache-2.0); Qwopus v2 fine-tune + BF16 GGUF by Jackrong (Apache-2.0); q27 repack by signalnine.