---

license: apache-2.0

pipeline_tag: text-generation

tags:

• merge
• slerp
• qwen3_5
• reasoning
• text-generation-inference
• gguf
• deepseek
• ornith

language:

• zh
• en

base_model: Qwen/Qwen3.5-9B

---

Qwen3.5-Ornith15-DSV4Pro-SLERP-9B

<div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;border-left:5px solid #7c3aed;background:#f5f3ff;padding:14px 18px;border-radius:0 10px 10px 0;margin-bottom:22px">

<b>Qwen3.5-Ornith15-DSV4Pro-SLERP-9B</b> is a tensor-wise SLERP merge (t=0.5) of two excellent post-trained models based on Qwen3.5-9B:

<a href="https://huggingface.co/ornith-ai/Ornith-1.5-9B"><b>ornith-ai/Ornith-1.5-9B</b></a> (MIT, agent/programming self-improvement) and

<a href="https://huggingface.co/Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B"><b>Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B</b></a> (Apache-2.0, DeepSeek-V4-Pro math/STEM distillation).

On a 561-item multi-dimensional benchmark, the merged model <b>outperforms both parent models</b>.

</div>

中文

Merge Details

• Method: Tensor-wise flattened SLERP at t=0.5 (consistent with mergekit's SLERP implementation), computed in float32 then converted to BF16 via round-to-nearest-even. Near-colinear tensors (|cos| > 0.9995) fall back to linear interpolation for numerical stability.
• Architecture: Qwen3.5 dense 9B — 32 layers of hybrid linear attention / full attention, 262K context.
• Structure source: Ornith-1.5-9B (tokenizer, chat template, shard layout). Optional MRP tensors unique to Jackrong are excluded since no Ornith counterpart exists; no single-side module is forcibly grafted into the result.
• Both original inputs are opened read-only; output is structure-validated and SHA-256 checksums are computed per shard (see merge_manifest.json).

Files

| File | Quantization | Size |

|---|---|---|

| Qwen3.5-Ornith15-DSV4Pro-SLERP-9B-IQ4_NL.gguf | IQ4_NL (~4.6 bpw) | ~5.1 GB |

Evaluation

All three models were evaluated under identical conditions using ZxBench v2026-08 (561 items, 10 dimensions), with each model run 3 times independently (1,683 scored items per run), averaged.

Evaluation Configuration

| Item | Value |

|---|---|

| Benchmark | ZxBench v2026-08 · 561 items · 10 dimensions |

| Runs | 3 independent full runs per model, averaged |

| Inference engine | llama.cpp b10516 (llama-server, OpenAI-compatible API) |

| Hardware | Tesla V100-SXM2 16GB, all layers offloaded to GPU (-ngl 99) |

| Context / Serving | Total context 294,912, --parallel 4 slots (~73K per item) |

| KV Cache | -ctk q4_0 -ctv q4_0 (Q4_0 quantized K/V cache) |

| Chat template | --jinja (thinking chain separated into reasoning_content field) |

| Sampling | temperature 0.6 · top-p 0.95 · top-k 20 · seed fixed at 42 |

| Output budget | max_tokens = 65,536 per item, reasoning mode enabled |

| Time limit | Hard 1,500-second limit per item; exceeded items are flagged, not scored 0 |

| Scoring | ZxBench deterministic scoring pipeline only (rule-based scoring + sandbox/hidden test execution verification); no AI Judge |

The three models tested under identical conditions:

• Qwen3.5-Ornith15-DSV4Pro-SLERP-9B — this merged model, IQ4_NL GGUF (~4.6 bpw)
• Ornith-1.5-9B — official release, Q4_K_M GGUF
• Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B — official MRP version, Q4_K_M GGUF

Notes — Please Read Before Comparing

> - Limited runs: Each model ran only 3 times, sufficient to show the ranking (merged model's three runs vary by ±0.32), but confidence is limited; score gaps of about 1 point between models should be treated as noise.

> - Output token budget: The 65,536 limit is generous but not infinite. A few heavy programming/agent items exhaust the budget or hit the time limit, and their answers are scored only on partial output, which may slightly underestimate all three models (especially for long programming items).

> - No AI Judge: Semantic scoring relies on deterministic rules and execution evidence weighted by coverage; purely semantic quality dimensions are not measured.

> - Different quantization per row: The merged model is IQ4_NL; both parent models are Q4_K_M. This comparison reflects differences in "deployable artifacts" rather than quantization-equivalent baselines.

> - All models score low on the tool/CLI workflow dimension (<40), indicating the criterion is stringent rather than any model being particularly weak.

Overall Results

| Model | Average Score | Pass Rate |

|---|---|---|

| Qwen3.5-Ornith15-DSV4Pro-SLERP-9B (IQ4_NL) | 68.41 | 69.0% |

| Ornith-1.5-9B (Q4_K_M) | 65.09 | 64.1% |

| Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B (Q4_K_M) | 63.38 | 63.9% |

Merged model run stability: 68.80 / 68.21 / 68.22 (±0.32).

Per-Dimension Results

| Dimension | Merged (IQ4_NL) | Ornith 1.5 (Q4_K_M) | DSV4-Pro (Q4_K_M) |

|---|---|---|---|

| Deep CLI Tasks | 88.33 | 79.83 | 77.37 |

| Instruction Following | 81.16 | 73.83 | 78.03 |

| Agent Workflow | 79.57 | 76.80 | 73.32 |

| Data Extraction | 79.60 | 76.78 | 77.19 |

| Hallucination Resistance | 72.91 | 74.27 | 68.12 |

| Reasoning & Math | 68.96 | 57.75 | 73.16 |

| Structured Output | 67.37 | 66.05 | 66.01 |

| Safety & Permissions | 63.86 | 60.09 | 60.99 |

| Coding | 60.78 | 58.42 | 52.25 |

| Tool/CLI/Workflow | 39.61 | 39.05 | 37.29 |

Key Takeaways

• The merged model leads the stronger parent (Ornith) by +3.32 points and the other parent by +5.03 points. The largest gains are in Deep CLI Tasks (+8.50) and Instruction Following (+7.33) — confirming that SLERP successfully fused complementary strengths from both parents (Ornith's agent tool-calling ability + DeepSeek's distilled structured reasoning).
• The parent models only lead on hallucination resistance (Ornith 74.27 vs 72.91).
• Reasoning & Math sits between the two parents (68.96): significantly above Ornith (+11.21) but below math-focused Jackrong (−4.20).

Usage (llama.cpp)

llama-server -m Qwen3.5-Ornith15-DSV4Pro-SLERP-9B-IQ4_NL.gguf \
  -c 131072 -ngl 99 --jinja --temp 0.6 --top-p 0.95 --top-k 20 \
  --host 127.0.0.1 --port 8765

This is a reasoning model: when started with --jinja, the thinking process </think>…</think> at the beginning of the response is returned separately in the reasoning_content field of the OpenAI-compatible API.

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8765/v1", api_key="EMPTY")
response = client.chat.completions.create(
    model="qwen3.5-merge",
    messages=[{"role": "user", "content": "Explain the Monty Hall problem."}],
    temperature=0.6, top_p=0.95,
)
print("Thinking:", getattr(response.choices[0].message, "reasoning_content", None))
print("Answer:", response.choices[0].message.content)

Recommended sampling parameters: temperature=0.6, top_p=0.95, top_k=20. Exact encoding tasks work equally well; avoid greedy decoding (reasoning models have a repetition risk).

Sources & License

• Common ancestor: Qwen/Qwen3.5-9B — both parent models are post-trained variants of this base (merged with equal weight, so no single parent is specified as the base model).
• Structure source: ornith-ai/Ornith-1.5-9B (MIT)
• Other source: Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B (Apache-2.0)
• The merged model is also available on HuggingFace: Dalvlad/Qwen3.5-Ornith15-DSV4Pro-SLERP-9B
• Original weights (full-precision BF16 and other variants) are hosted on ModelScope
• Released under Apache-2.0; the respective licenses of both source checkpoints have been respected. Model merging is experimental; neither parent model's baseline performance nor safety guarantees are preserved — please evaluate before use.

Acknowledgments

• Evaluation uses ZxBench — a localized large model benchmark system (561 items, 10 dimensions, deterministic scoring). Thank you to the authors for open-sourcing this evaluation framework.
• Inference powered by llama.cpp; merging uses a lightweight NumPy SLERP implementation equivalent to mergekit.

Limitations

• Evaluation results are based on a single benchmark framework (ZxBench, deterministic scoring, no LLM Judge); absolute scores are not comparable across different evaluation systems.
• GGUF is a text-only export of this multimodal architecture family (without the vision tower).
• Subject to common limitations of 9B reasoning models: hallucination, unstable long-range planning, sensitivity to prompt formatting.

Citation

@misc{qwen35_ornith_dsv4pro_slerp_9b,
  title  = {Qwen3.5-Ornith15-DSV4Pro-SLERP-9B},
  author = {Local merge, tensor-wise flattened SLERP t=0.5},
  year   = {2026},
  note   = {Merged from ornith-ai/Ornith-1.5-9B and Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B}
}