Qwen3.6-27B family
Evaluation comparison — seven builds, one methodology
What this covers
- Four bf16 27B builds side by side — base, Tess-4 tune, Aeon tune, ThinkingCap tune.
- Both measurements of the Fable-Fusion-711 merge: its unpacked GGUF Q8_0 weights, and the native bf16 weights that were never quantized (new on 2026-07-25).
- Qwen3.6-35B-A3B (MoE) included as a reference point.
- Every build measured with the same methodology against its own live vLLM endpoint (tcclaviger/vllm:latest), July 2026.
- Every build served with YaRN rope expansion (factor 2.0, original_max_position_embeddings 262144) at a
KV/context length of
--max-model-len 524288, TP4, MTP k=5. This matters for repeatability: throughput and long-context recall both move with context length and rope config. One exception — the Fusion-711 Q8 run was served at 524188, a 100-token typo in that compose.
Credits & thanks
These are other people's models. All I did was measure them. Thanks to everyone below for building and releasing this work.
- Qwen3.6-27B & Qwen3.6-35B-A3B — the Qwen Team, Alibaba.
- Tess-4-27B — Migel Tissera (migtissera).
- ThinkingCap-Qwen3.6-27B — BottleCap AI (bottlecapai).
- Qwen3.6-27B-AEON-Ultimate-Uncensored — AEON-7 (pipeline, configuration, testing), crediting FernflowerAI (SSM conv1d outlier repair), Wangzhang Wu (abliterix v1.4), Philipp Emanuel Weidmann (heretic, upstream of abliterix), and the abliteration literature — Arditi et al. (2024), grimjim (2025), Huang et al. (2025), Xie et al. (2026) — on top of Qwen3.6-27B.
- Fable-Fusion-711 — a collaboration credited by DavidAU (fine tunes, multi-stage) with Nightmedia (merge & benching), TeichAI (Polaris dataset), armand0e (Fable-5 traces) and trohrbaugh (heretic'ing), on top of Qwen3.6-27B.
WikiText-2 perplexity
Tool-eval final score
Accuracy benches
Multiple-choice accuracy
Logprobs accuracy — raw acc
Codeneedle positional recall
Tool-eval Hard Mode (Category P) — thinking-ON
Decode throughput — ISL 128
Decode throughput — ISL 512
Prefill speed — ISL 128
Prefill speed — ISL 512
All numbers
All builds serve on the tcclaviger/vllm:latest image. Base Qwen3.6-27B figures are the 2026-07-12 re-measurement; Tess/Aeon/ThinkingCap/35B are their 2026-07-11/12 runs; both 711 runs are 2026-07-25.
Credits: codeneedle (positional recall) originally by Alexander Ziskind, expanded test suite by tcclaviger (Rob Smith). Tool-calling scenarios run on tool-eval-bench by SeraphimSerapis (Tim Messerschmidt), scenario methodology adapted from ToolCall-15 by stevibe. GSM8K, MMLU, IFEval via tool-eval-bench built-in benches at their defaults.