library_name: transformers
license: apache-2.0
tags:
- merge
- qwen3_5
- qwen3.5
- reasoning
- coding
- agentic
- conversational
- long-context
- transformers
- heretic
- uncensored
- decensored
- abliterated
- reproducible
base_model: xtools-at/Qwen3.5-9B-Claude-3xMerge
datasets: - armand0e/claude-fable-5-claude-code
- lambda/hermes-agent-reasoning-traces
- Jackrong/Claude-opus-4.7-TraceInversion-5000x
- Jackrong/Claude-opus-4.6-TraceInversion-9000x
pipeline_tag: text-generation
This is a decensored version of xtools-at/Qwen3.5-9B-Claude-3xMerge, made using Heretic v2.0.0.dev0
[!TIP]
This model is reproducible!See the README in the
reproducedirectory for more information.
Arbitrary-Rank Ablation (ARA) parameters
| Parameter | Value |
|---|---|
| start_layer_index | 16 |
| end_layer_index | 23 |
| preserve_good_behavior_weight | 0.6496 |
| steer_bad_behavior_weight | 0.0006 |
| overcorrect_relative_weight | 0.9719 |
| neighbor_count | 14 |
Performance
| Metric | This model | Original model (xtools-at/Qwen3.5-9B-Claude-3xMerge) |
|---|---|---|
| Refusals | 1/100 | 99/100 |
| KL divergence | 0.0205 | 0 (by definition) |
Qwen3.5-9B Claude Mythos/Fable/Opus Triple-Merge
A weighted merge of three Qwen3.5-9B fine-tunes, combining reasoning-oriented generation, conversational behavior, and coding-focused agent workflows.
This experimental merge brings together three complementary model variants:
- Qwythos-9B-v2: Reasoning-oriented generation, reduced repetitive output, and long-context support.
- Qwen3.5-9B-Fable-5-v1: Fine-tuning on Claude Fable 5 / Claude Code traces, with reported improvements on selected language-understanding benchmarks.
- Qwopus3.5-9B-Coder: Coding, structured reasoning, function calling, and agent-oriented tool-use workflows.
The objective is to combine general-purpose reasoning with conversational generation and software-development capabilities in a single approximately 9B-parameter model.
This is an experimental checkpoint merge. Its performance, reliability, and compatibility must be evaluated independently of the source models.
Model overview
| Property | Details |
|---|---|
| Base architecture | Qwen3.5-9B |
| Model type | Weighted checkpoint merge |
| Merge method | Linear weighted average |
| Primary focus | Reasoning, conversational generation, and coding |
| Additional focus | Agent workflows, tool use, and long-context tasks |
| Parameter class | Approximately 9B |
| Declared license | Apache-2.0; verify upstream terms before redistribution |
Merge recipe
The merged checkpoint combines three source models using the following configured coefficients.
| Source model | Weight | Primary emphasis |
|---|---|---|
| empero-ai/Qwythos-9B-v2 | 1.0 | Mythos reasoning, output stability, and reduced repetition |
| TeichAI/Qwen3.5-9B-Fable-5-v1 | 0.8 | Claude Fable 5 / Claude Code trace fine-tuning |
| Jackrong/Qwopus3.5-9B-Coder | 0.6 | Opus coding, reasoning, and agentic tool use |
Merge method: Linear weighted average.
Source models
1. Qwythos-9B-v2
Qwythos-9B-v2 is an Empero AI fine-tune built on the Qwen3.5-9B architecture and the earlier Qwythos-9B-Claude-Mythos-5-1M model.
The model emphasizes reasoning-oriented generation and long-context use. Its v2 release specifically targets repetitive output and degeneration during greedy or low-temperature decoding.
The author describes using Final-Token Preference Optimization (FTPO) to identify the token that begins a repetition loop and train the model to prefer coherent alternatives. The published training procedure reports approximately 2,000 automatically mined preference tuples.
The v2 release also restores the base model's multi-token-prediction (MTP) head and updates model behavior to reduce unwanted self-identification in unrelated responses.
The source card reports a YaRN-scaled context length of approximately 1,048,576 tokens, although this setting should not be assumed to transfer unchanged to the merged checkpoint.
Role in this merge: Provides the primary reasoning-oriented foundation, with emphasis on stable generation and reducing repetitive output.
Important distinction: Qwythos v2 is described as a robustness and hygiene update rather than a major increase in raw capabilities. Its reported benchmarks belong to the source model, not this merge.
2. Qwen3.5-9B-Fable-5-v1
TeichAI's Fable 5 is a Qwen3.5-9B fine-tune trained on the publicly identified armand0e/claude-fable-5-claude-code dataset.
The author describes extracting, formatting, and masking the training data using Teich, with training accelerated by Unsloth and Hugging Face TRL.
The source card reports the following results:
| Benchmark | Fable-5-v1 | Qwen3.5-9B |
|---|---|---|
| ARC | 0.624 | 0.553 |
| ARC-Easy | 0.806 | 0.712 |
| BoolQ | 0.891 | 0.892 |
These are the source model's reported results and should not be attributed to the merged checkpoint.
The dataset name indicates Claude Fable 5 / Claude Code traces. The source card does not provide sufficient evidence to characterize every training example or establish that the model is exclusively a creative-writing fine-tune.
Role in this merge: Adds a distinct conversational and instruction-tuning trajectory learned from the Fable 5 dataset.
3. Qwopus3.5-9B-Coder
Qwopus3.5-9B-Coder is Jackrong's coding-oriented Qwen3.5-9B fine-tune. Its published tags and model documentation emphasize software development, structured reasoning, function calling, and agentic workflows.
The model tree identifies Qwen3.5-9B as its underlying architecture and lists Qwopus3.5-9B-v3.5 in its fine-tuning lineage.
The publicly listed training datasets are:
lambda/hermes-agent-reasoning-tracesJackrong/Claude-opus-4.7-TraceInversion-5000xJackrong/Claude-opus-4.6-TraceInversion-9000x
The training approach combines agent-oriented reasoning traces with Claude Opus-style trace-inversion data. This provides a training emphasis on structured problem-solving and coding-related agent workflows.
Role in this merge: Adds coding-oriented behavior, tool-use patterns, and agentic reasoning.
Base model
Qwen3.5-9B
Qwen3.5-9B is the shared underlying architecture for all three source models.
The official model card describes a 9B-parameter causal language model with a vision encoder. Its hybrid architecture combines gated linear-attention layers with full-attention layers, supporting long-context processing while retaining multimodal capabilities.
The base model documents a native context length of 262,144 tokens, with extended context supported through appropriate RoPE scaling configurations.
These are base-model specifications. The merged checkpoint may differ in context behavior, multimodal compatibility, MTP support, or inference requirements.
Training data and provenance
The merged checkpoint is produced by combining three independently fine-tuned models. It was not directly trained on a newly assembled dataset containing all of the samples listed below.
Fable 5 dataset
| Dataset | Description |
|---|---|
| armand0e/claude-fable-5-claude-code | Claude Fable 5 / Claude Code traces used for the Fable 5 fine-tune |
The source model card describes the extraction, formatting, and masking of the data for training. It does not fully document every preprocessing detail or the provenance of every underlying trace.
Qwopus Coder datasets
| Dataset | Intended role |
|---|---|
| lambda/hermes-agent-reasoning-traces | Agent-oriented reasoning and tool-use traces |
| Jackrong/Claude-opus-4.7-TraceInversion-5000x | Claude Opus 4.7-style trace-inversion data |
| Jackrong/Claude-opus-4.6-TraceInversion-9000x | Claude Opus 4.6-style trace-inversion data |
These datasets are listed on the Qwopus Coder source card. Their presence does not imply that all samples were used in equal proportions or that the same data was used by the other source models.
Qwythos training data
Qwythos-9B-v2's documented v2 training stage uses approximately 2,000 automatically mined preference tuples for FTPO-based repetition correction. These examples are constructed from the model's own looping behavior and coherent alternative tokens.
The source model's broader training lineage includes empero-ai/Qwythos-9B-Claude-Mythos-5-1M. The Qwythos v2 card does not identify a complete public dataset manifest for the earlier training stages, so those datasets are not listed here as verified public dataset IDs.
Provenance limitations
Dataset names and training details above are based on the upstream model cards. They should not be interpreted as a complete record of every preprocessing stage, sample count, or training mixture.
Consult the source model and dataset cards for their latest descriptions, licenses, and acknowledgements.
Intended use
This merge is intended for experimentation with a compact model combining general reasoning, conversational generation, and coding-oriented behavior.
Potential use cases include:
- General reasoning: Analytical questions, logical deduction, and multi-step problem-solving.
- Programming: Code generation, debugging, code explanation, and software-development assistance.
- Agent workflows: Experimentation with function calling, tool-assisted tasks, and iterative problem-solving.
- Conversational generation: Dialogue, instruction following, and general-purpose text generation.
- Long-context experimentation: Tasks involving large prompts, provided the merged model and inference engine support the desired context length.
These are intended use cases inferred from the source models, not independently verified performance guarantees.
Reasoning and generation behavior
The source models have different training histories and generation characteristics.
- Qwythos v2 targets repetitive output and reasoning-oriented generation.
- Fable 5 contributes behavior learned from its Claude Fable 5 / Claude Code traces.
- Qwopus Coder emphasizes structured reasoning, programming, and tool-oriented workflows.
Merging these models may preserve some of these characteristics, alter others, or introduce new failure modes. In particular, the merge may exhibit repetition, inconsistent formatting, unexpected reasoning behavior, or unreliable tool calls.
The presence of reasoning markers or structured tool-call outputs does not guarantee that the model's reasoning is correct or that its tools will execute successfully.
Limitations
- Merge interference: Independently fine-tuned checkpoints can contain conflicting parameter updates. A weighted merge does not guarantee that every capability of every source model is retained.
- Unverified merged performance: Source-model benchmark results cannot be attributed to this merged checkpoint without direct evaluation.
- Coding reliability: Generated code may contain bugs, security vulnerabilities, or incorrect API usage. Test generated code in an appropriate isolated environment.
- Tool-use reliability: Function calling depends on the tokenizer, chat template, inference engine, tool definitions, and execution environment.
- Repetitive output: Qwythos v2 specifically targets looping behavior, but its reported improvements do not establish that this merge is free from repetition.
- Long-context behavior: The Qwythos v2 context setting may not remain compatible with the merged model. Validate the actual configuration and inference engine before using extended contexts.
- Multimodal compatibility: The base architecture supports image inputs, but this merge's visual functionality depends on the retained model components and processor files.
- MTP compatibility: Qwythos v2 restores an MTP head, but its compatibility with the merged weights and speculative-decoding implementations has not been established.
- Safety and factuality: The model may generate inaccurate, misleading, or inappropriate responses. It should not be treated as a verified source of truth.
Use independent verification and human review for consequential decisions.
Evaluation
No independent benchmark results are reported here for the merged checkpoint.
The source models publish their own evaluations, but differences in prompts, sampling settings, datasets, and scoring procedures make direct comparisons difficult. Their reported results should not be presented as results for this merge.
Recommended evaluation areas include:
| Area | Evaluation focus |
|---|---|
| General reasoning | Multi-step problems, logical deduction, and instruction following |
| Coding | Code generation, debugging, and software-engineering benchmarks |
| Agentic behavior | Tool selection, argument formatting, and multi-step tool execution |
| Generation stability | Repetition, looping, and low-temperature decoding |
| Conversational quality | Coherence, formatting, and adherence to instructions |
| Long context | Accuracy and consistency at increasing context lengths |
| Multimodal support | Image understanding, if the required components are retained |
For reproducible comparisons, document the exact model revisions, inference settings, prompts, quantization, context length, and evaluation methodology.
Acknowledgements
This merge builds on the work of:
- The Qwen team for the Qwen3.5 model family.
- Empero AI for Qwythos.
- TeichAI and the contributors to the Fable 5 dataset.
- Jackrong for the Qwopus Coder fine-tune and trace-inversion datasets.
- The authors and maintainers of the listed training datasets.
- The open-source fine-tuning and inference communities.
Please consult the original model cards for additional acknowledgements and citations.
License
This repository declares the Apache-2.0 license. Verify the applicable license and usage terms for the base model, all three source checkpoints, and the underlying datasets before redistribution or commercial use.