Qwen3.5-27B-Omnimerge-v2-GGUF

GGUF quantizations of ManniX-ITA/Qwen3.5-27B-Omnimerge-v2 — a 3-way weight-space merge using the Omnimerge v2 method (OBIM + DAREx + EMR).

V2 outperforms v1 across all benchmarks, with +8 pp on GPQA Diamond reasoning and +16 pp over the best source model.

Benchmark Results (Q6_K)

Benchmark	Omnimerge v1	Omnimerge v2	Best source (Claude-distill)
GPQA Diamond (198q, flex)	61.11%	69.19% (+8.08 pp)	53.03%
MBPP pass@1	71.80%	74.60% (+2.80 pp)	71.20%
HumanEval pass@1	79.88%	79.27% (-0.61 pp)	76.22%

Recommended Usage

llama-server -m Qwen3.5-27B-Omnimerge-v2-Q6_K.gguf -c 32768 -ngl 99 \
    --reasoning-format deepseek --reasoning-budget 16384 \
    --temp 0.6 --top-p 0.95 --top-k 20 --dry-multiplier 0.5

Source Models

Source	Weight
Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled	0.40
ValiantLabs/Qwen3.5-27B-Esper3.1	0.35
Jackrong/Qwen3.5-27B-Gemini-3.1-Pro-Reasoning-Distill	0.25

See the model card for full methodology.

License

Apache-2.0