This is a decensored version of xtools-at/Qwen3.5-4B-Claude-3xMerge-Jackrong, made using Heretic v2.0.0.dev0

[!TIP] This model is reproducible!

See the README in the reproduce directory for more information.

Arbitrary-Rank Ablation (ARA) parameters

Performance

---

Qwen3.5-4B Claude 3xMerge Jackrong

A weighted merge of three Qwen3.5-4B fine-tunes by Jackrong, combining Claude Opus-style reasoning distillation, coding-oriented fine-tuning, and structured multi-step reasoning.

The goal of this merge is to bring together complementary capabilities from the source models in a single compact checkpoint, with an emphasis on reasoning, programming, debugging, and agent-oriented workflows.

This is an experimental model merge. Its performance characteristics may differ from those of its individual source models.

Overview

• Base architecture: Qwen/Qwen3.5-4B
• Model type: Weighted checkpoint merge
• Merge method: Linear weighted average
• Primary focus: Reasoning, coding, and general problem-solving
• Additional focus: Structured reasoning and agentic tool-use workflows
• License: Apache-2.0, as declared for this release

Merge recipe

The merged checkpoint combines the following three fine-tunes using the specified weights.

The weights above are the configured merge coefficients, not percentages of training data or measured contributions to the final model. The exact normalization and implementation details depend on the merge procedure used.

Source models

1. Qwen3.5-4B-Claude-4.6-Opus-Reasoning-Distilled-v2

Model card

This fine-tune focuses on general-purpose reasoning distilled from Claude Opus 4.6-style traces. Its training emphasizes structured problem decomposition, mathematical and logical reasoning, and more economical reasoning patterns.

The source model aims to reduce repetitive reasoning and unnecessary intermediate steps while retaining the ability to work through complex problems.

Role in this merge: General reasoning, analytical problem-solving, and instruction-following behavior.

2. Qwopus3.5-4B-Coder

Model card

This coding-oriented fine-tune targets software development workflows, including code generation, debugging, structured reasoning, and tool-assisted problem-solving.

Its training approach incorporates trace inversion, coding-oriented supervision, and agent trajectories designed to teach tool-use behavior and iterative task execution.

Role in this merge: Programming, debugging, and coding-related agent workflows.

3. Qwopus3.5-4B-v3

Model card

This reasoning-focused fine-tune emphasizes the structure and clarity of multi-step reasoning. Its training approach aims to produce more explicit reasoning traces and improve the consistency of reasoning across tasks.

The source model also explores execution-driven refinement and tool-calling behavior for multi-step agent workflows.

Role in this merge: Explicit reasoning structure, complex problem decomposition, and complementary coding and agent-oriented behavior.

Base model

Qwen3.5-4B

Official model card

Qwen3.5-4B is the underlying model family for this merge. It provides the shared architecture and pretrained capabilities from which the three fine-tunes are derived.

The original model supports text and visual inputs. However, the capabilities preserved by this particular merge depend on the merged weights, model configuration, tokenizer, processor, and inference implementation. Multimodal functionality should be tested rather than assumed.

Training data and provenance

This checkpoint is produced by merging existing fine-tuned models rather than by directly training a new checkpoint on the combined datasets.

The source-model training data listed for this release includes:

These datasets are listed as provenance for the source fine-tunes. Their inclusion does not mean that every sample was used to train every source model, nor that the datasets were directly combined when producing this merge.

For detailed information about individual training runs, filtering procedures, and data composition, consult the linked source-model and dataset cards.

Intended use

This model is intended for experimentation with a compact reasoning-oriented model that combines general problem-solving and coding-related behavior.

Potential use cases include:

• General reasoning: Multi-step questions, logical deduction, and analytical tasks.
• Programming: Code generation, code explanation, and debugging.
• Technical problem-solving: Breaking complex tasks into smaller steps and evaluating possible solutions.
• Agentic workflows: Experimentation with structured tool calls and iterative execution in compatible environments.
• Local inference: Evaluation and deployment in environments that support the underlying Qwen3.5 architecture and the merged checkpoint's configuration.

These are intended use cases inferred from the source models, not guarantees of performance.

Limitations

• Merge behavior: Combining fine-tuned checkpoints can produce interference between learned behaviors. The resulting model may not retain every capability of each source model.
• Unverified performance: The merge should be evaluated independently. Results reported for the source models cannot be attributed directly to this checkpoint.
• Reasoning reliability: Generated reasoning can contain logical errors, unsupported assumptions, repetition, or incorrect conclusions.
• Coding reliability: Generated code may contain bugs, security vulnerabilities, or incorrect API usage. Test code in an appropriate isolated environment.
• Tool use: Tool-calling behavior depends on the chat template, inference engine, tool definitions, and execution environment. Compatibility should be verified before deployment.
• Multimodal compatibility: Although the base model supports visual inputs, this merge's visual capabilities and processor compatibility have not been independently established.
• Training-data limitations: The characteristics and limitations of the source models and their training datasets may carry over to the merged checkpoint.
• Safety: This is an experimental community merge and has not been independently validated for safety-critical or high-stakes applications.

Do not rely on this model as the sole source of truth for medical, legal, financial, security-critical, or other high-impact decisions.

Evaluation

No independent benchmark results are reported here for this merged checkpoint.

For a meaningful evaluation, compare it against the base model and all three source models using the same prompts, inference settings, context limits, and evaluation datasets. Coding, mathematical reasoning, instruction following, and tool-use tasks are useful starting points.

Benchmark results should be added only after testing the merged checkpoint directly, with the evaluation setup and scoring methodology documented.

Acknowledgements

This merge builds on the work of the Qwen team and Jackrong, along with the contributors and dataset authors behind the linked fine-tunes and training datasets.

Please refer to the original model cards for their respective training details, acknowledgements, and citations.

License

This repository declares the Apache-2.0 license. Consult the source-model and dataset licenses and terms as well, particularly before redistribution or commercial use.