A compact model combining mathematical and scientific reasoning with coding, debugging, and agentic capabilities. This merge brings together two complementary Qwen3.5 fine-tunes, with an emphasis on structured problem-solving, practical coding skills, and fewer unnecessary refusals.
[!NOTE] Looking for a smaller model? Try the 4B variant.
The goal was to combine two complementary strengths into one lightweight model:
The model was created through weight-space merging using SLERP (Spherical Linear Interpolation), rather than additional fine-tuning.
| Setting | Value |
|---|---|
| Merge method | SLERP |
| Interpolation parameter | t = 0.4 |
| First model (higher weight) | xtools-at/Qwopus3.5-9B-Coder-UNCENSORED-heretic |
| Second model | xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic |
With t = 0.4, the first parent receives the higher interpolation weight, giving the merge a stronger coding and agent-oriented foundation while incorporating the reasoning strengths of the second model.
Model repository · Original Jackrong fine-tune
A Qwen3.5-based coding and agent model focused on practical programming workflows, debugging, bug localization, code patches, tool routing, and function calling.
Its training combines reconstructed Claude reasoning traces with agent trajectories from real-world coding and tool-use tasks.
Model repository · Original Jackrong fine-tune
A Qwen3.5-based reasoning model distilled from DeepSeek-V4-Pro in Max Effect mode. Its fine-tuning uses approximately 250K mathematics and STEM samples, focusing on structured derivations, problem decomposition, and scientific reasoning.
Both parent models are Heretic-based decensored variants. The merged checkpoint has also been tested, and its abliterated behavior has been confirmed.
The parent models draw on different training-data sources:
Mathematics & STEM
The DeepSeek-based parent uses approximately 250K mathematics and STEM examples distilled from DeepSeek-V4-Pro. Its model card does not identify a separate public dataset repository.
Claude Trace Inversion
The Qwopus parent references two reconstructed reasoning datasets:
Roman1111111/claude-opus-4.6-10000x.lordx64/reasoning-distill-claude-opus-4-7-max.These datasets use Trace Inversion to reconstruct detailed synthetic reasoning traces from problems, answers, and compressed summaries. They are reconstructed training data, not direct access to private internal reasoning traces.
Research reference: How to Steal Reasoning Without Reasoning Traces.
Agent & Tool-Use Data
lambda/hermes-agent-reasoning-traces provides multi-turn agent trajectories originating from Kimi-K2.5 and GLM-5.1-FP8. The examples cover tool calling, terminal operations, coding, repository work, and browser automation.
This model is intended for experimentation and local use, particularly:
As a compact class model, it aims to provide a useful balance between capability and local deployment requirements.
Due to resource constraints, this initial evaluation was limited to the 4B variant, comparing it to the upstream parent models, using a small set of benchmarks. The results suggest that fine-tuning has largely preserved general benchmark performance, with no clear indication of overall degradation. Scores shifted slightly across tasks, with this model performing similarly to its parents on ARC-Easy and doing marginally better on the ARC-Challenge benchmark, and closing the gap between the parents on GSM8K.
This is a preliminary assessment rather than a comprehensive evaluation. Testing may be expanded in the future, including runs on the 9B model and additional specialized STEM, SWE and Terminal benchmarks.
Best score per task is in bold. All models loaded with BNB 4-bit quantization. Full test splits were used for all tasks.
| Task (metric) | Jackrong/DeepSeek-V4-Pro-Qwen3.5-4B | Jackrong/Qwopus3.5-4B-Coder | xtools-at/Qwen3.5-4B-DeepThought-STEM | xtools-at/Wukong-4B |
|---|---|---|---|---|
| ARC-Easy (acc_norm) | 0.7567 | 0.7508 | 0.7778 | 0.7525 |
| ARC-Challenge (acc_norm) | 0.5350 | 0.5367 | 0.5478 | 0.5512 |
| GSM8K (exact_match) | 0.7756 | 0.7384 | 0.7400 | 0.7650 |
This is an experimental community merge, not a newly trained or comprehensively evaluated foundation model.
Benchmark results reported by the parent models should not be interpreted as results for this merged checkpoint. Merging can affect the performance and behavior of both parents, and the resulting model has not been independently benchmarked.
The decensored and abliterated behavior also means that refusal patterns and safety boundaries may differ from conventional instruction-tuned models. Generated code and tool calls should be reviewed before use in real environments.
Both immediate parent model repositories declare Apache-2.0 licensing. The referenced datasets also list Apache-2.0 in their Hugging Face metadata. Please review the original repositories and applicable third-party data terms when redistributing or using the model commercially.
Full credit goes to the original authors and contributors:
Thanks to everyone involved in making these models, datasets, and tools available to the community. ❤️
model-00002-of-00005.safetensors 4.3 GBmodel-00004-of-00005.safetensors 4.3 GBmodel-00001-of-00005.safetensors 4.3 GBmodel-00003-of-00005.safetensors 4.3 GBmodel-00005-of-00005.safetensors 2.0 GBQwen/Qwen3.5-9B