Qwen3.5-9B Claude Mythos/Fable/Opus 5xMerge

A weighted merge of five Qwen3.5-9B fine-tunes combining Claude Fable 5 agentic coding traces, Claude Mythos reasoning distillation, and Claude Opus 4.6-style reasoning fine-tuning.

The goal is to combine several complementary post-training approaches in one checkpoint:

• Agentic coding: Multi-turn software-development tasks, terminal workflows, and tool-use traces.
• Self-distilled coding behavior: On-policy self-distillation from Claude Fable 5 coding-agent trajectories.
• General reasoning: Structured problem-solving across mathematics, science, and analytical tasks.
• Reasoning style: Claude-inspired thinking patterns and explicit intermediate reasoning.
• Conversational generation: General-purpose dialogue, instruction following, and extended text generation.
• Long-context experimentation: A model lineage that includes a Qwythos variant configured for an extended context window.

This is an experimental weighted merge. Its performance and behavior may differ substantially from the individual source models, and none of their benchmark results should be assumed to transfer directly to this checkpoint.

Model overview

Merge recipe

The merge combines five independently fine-tuned checkpoints using the following configured coefficients.

Merge method: Linear weighted average.

Source models

1. Qwable-9B-Claude-Fable-5

Source model card

Qwable is a full-parameter supervised fine-tune of Qwen3.5-9B developed by Empero. It uses a curated mixture of agentic coding and reasoning traces generated by Claude Fable 5 and a GPT-5.5 terminal agent.

The source card reports two training datasets:

The training data is heavily weighted toward Fable 5 traces, which make up approximately 97% of the reported examples. The training procedure uses assistant-only loss, long sequences, and full-parameter fine-tuning.

The source card reports held-out validation loss and qualitative generation tests, but notes that a full coding, agentic, and safety benchmark suite is still in progress.

Role in this merge: Provides the primary contribution for agentic coding, terminal-oriented workflows, and multi-turn reasoning traces.

2. Qwen3.5-9B-Fable-5-SDFT

Source model card

This model applies Self-Distillation Fine-Tuning (SDFT) to Qwen3.5-9B using coding-agent and tool-use trajectories from the Fable 5 dataset.

Unlike conventional supervised fine-tuning, SDFT trains on responses sampled from the current model and uses teacher-conditioned distributions to guide the model toward expert behavior. The documented training procedure uses reverse-KL distillation and on-policy rollouts.

The source card reports:

• 2,693 filtered examples derived from armand0e/claude-fable-5-claude-code.
• A target context length of 65,536 tokens.
• A maximum prompt length of 57,344 tokens.
• Up to 8,192 new tokens per training rollout.
• 600 optimizer steps using a LoRA-based training procedure, with the final adapter merged into BF16 weights.

The source card explicitly notes that private/internal reasoning fields are not included in the teacher reference.

Role in this merge: Adds an alternative training trajectory over related Fable 5 coding-agent data, emphasizing on-policy response generation and tool-use behavior.

Because this source shares training-data lineage with Qwable, the two models should not be treated as independent sources of completely distinct knowledge.

3. Qwythos-9B-Claude-Mythos-5-1M

Source model card

Qwythos is a full-parameter reasoning fine-tune of Qwen3.5-9B developed by Empero. It is post-trained on more than 500 million tokens of Claude Mythos and Claude Fable traces, supplemented by structured reasoning traces generated with Empero's internal rethink tool.

The documented training domains include:

• Mathematics and general problem-solving.
• Programming and code-related reasoning.
• Scientific and biomedical analysis.
• Agentic tool use and multi-turn tasks.

The source model is designed to produce structured thinking blocks followed by final answers. Its author also reports native function-calling behavior and an extended context configuration using YaRN.

The source card specifies a 1,048,576-token context configuration, extending the base architecture's native context length through RoPE scaling. This is a source-model configuration, not a guaranteed capability of the merged checkpoint.

The complete underlying dataset manifest is not publicly enumerated in the source card. In-house generated reasoning traces and the described Claude-derived training mixture should therefore be treated as provenance information rather than a fully reproducible public dataset list.

Role in this merge: Contributes general reasoning, scientific problem-solving, structured thinking behavior, and long-context training lineage.

4. Claude-4.6-OS Auto-Variable THINKING

Source model card

This is an Unsloth fine-tune of Qwen3.5-9B using a Claude 4.6 OS dataset described by DavidAU as containing four Claude-related datasets.

The author describes the training as relatively mild, with a focus on adjusting reasoning behavior and thinking-block structure while attempting to preserve the base model's existing capabilities.

The source card reports image-input testing as successful and publishes selected benchmark comparisons against Qwen3.5-9B. These are author-reported results for the source model only.

The four individual dataset repository IDs are not enumerated in the source card, so they cannot be reliably listed here.

Role in this merge: Adds another Claude Opus 4.6-style reasoning fine-tuning trajectory, with an emphasis on reasoning structure and generation behavior.

5. Claude-4.6-HighIQ THINKING

Source model card

This is a second DavidAU fine-tune of Qwen3.5-9B using a large Claude 4.6 distillation dataset.

The author describes the training as designed to improve thinking generation and reasoning-oriented benchmark performance while attempting to preserve the capabilities of the base model.

The source card includes results for several knowledge, reasoning, and language-understanding benchmarks. These results are specific to the source checkpoint and its evaluation setup.

As with the OS Auto-Variable variant, the source card does not provide a complete public manifest identifying all individual training datasets.

Role in this merge: Adds a complementary Claude 4.6-derived fine-tuning trajectory focused on reasoning and structured thinking.

The two DavidAU checkpoints share a similar training lineage. Their coefficients should not be interpreted as evidence that they contribute completely independent capabilities.

Base model

Qwen3.5-9B

Official model card

Qwen3.5-9B is the shared architecture underlying all five source models.

The Qwen3.5 family supports language and visual understanding, reasoning, coding, and agent-oriented tasks. The architecture combines gated linear-attention layers with full-attention layers, enabling efficient processing of long sequences.

The official base model documents a native context length of 262,144 tokens, with extended contexts possible under compatible configurations.

These are base-model specifications. They do not guarantee that every capability, context setting, or model component is preserved by the merge.

Training data and provenance

The merged checkpoint is produced by combining five independently fine-tuned models. It was not directly trained on a new dataset containing all of the examples described below.

Publicly identified datasets

The SDFT model uses a filtered subset of the Fable 5 dataset, so its training examples overlap in lineage with the dataset used by Qwable.

Additional source-model training data

Qwythos: More than 500 million tokens of Claude Mythos and Claude Fable traces, combined with in-house reasoning traces generated using Empero's rethink tool. The full public dataset manifest is not provided.

DavidAU Claude-4.6-OS Auto-Variable: Four Claude-related datasets, as described by the author. Individual public dataset IDs are not enumerated.

DavidAU Claude-4.6-HighIQ: A large Claude 4.6 distillation dataset. Individual public dataset IDs are not enumerated.

These descriptions reflect the source authors' published documentation. They do not establish the exact overlap between datasets, the sample count used by each training run, or the relative proportion of each source's data in the merged checkpoint.

For full provenance, consult the linked source-model and dataset cards.

Intended use

This merge is intended for experimentation with a compact model combining reasoning-oriented generation, conversational behavior, and coding-agent capabilities.

Potential use cases include:

• Software development: Code generation, debugging, code explanation, and multi-step implementation tasks.
• Terminal and agent workflows: Structured tool calls, command-line task planning, and iterative task execution in compatible environments.
• General reasoning: Analytical questions, mathematics, scientific reasoning, and structured problem-solving.
• Conversational generation: Dialogue, instruction following, and general-purpose text generation.
• Long-context experimentation: Tasks involving large documents, codebases, or extended conversations, provided the actual merged configuration supports the required context length.
• Creative writing: Narrative generation and extended prose, subject to evaluation of the merged model's actual behavior.

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

Reasoning and tool-use behavior

The source models differ in their fine-tuning methods and generation styles.

• Qwable emphasizes supervised learning from long-form agentic coding and reasoning traces.
• Fable 5 SDFT uses on-policy self-distillation on coding-agent trajectories.
• Qwythos emphasizes structured reasoning and multi-domain problem-solving.
• The DavidAU models emphasize Claude Opus 4.6-style reasoning and modified thinking patterns.

The merge may preserve some of these characteristics while weakening or altering others. It may also exhibit repetitive reasoning, inconsistent formatting, incorrect tool calls, or unexpected shifts between reasoning styles.

A visible <think> block does not guarantee correctness. Tool-call syntax likewise does not guarantee that a tool exists, that the arguments are valid, or that the requested operation is safe.

Limitations

• Merge interference: Independently fine-tuned checkpoints can contain conflicting parameter updates. A weighted merge does not guarantee retention of all source-model capabilities.
• Overlapping training data: Qwable and Fable 5 SDFT share Fable 5 data lineage, while the DavidAU models share a Claude 4.6 training lineage. The merge combines related training trajectories rather than five wholly independent knowledge sources.
• Unverified performance: No independent benchmark results are reported for this merged checkpoint.
• Coding reliability: Generated code can contain bugs, insecure patterns, or incorrect API usage. Execute code only in an appropriately isolated environment.
• Tool-use reliability: Function calling depends on the tokenizer, chat template, inference engine, tool definitions, and execution environment.
• Reasoning reliability: Generated explanations and intermediate steps may be incorrect, repetitive, or internally inconsistent.
• Long-context limitations: Qwythos's extended context configuration does not guarantee that this merge supports a 1M-token context. Validate the actual configuration and inference engine.
• Multimodal compatibility: The Qwen3.5-9B base supports image inputs, but the merged checkpoint's visual functionality depends on retaining compatible weights, configuration, and processor files. Some source training procedures adapted only the language pathway.
• Safety and factuality: The model can produce misleading, harmful, or fabricated content. It has not been independently validated for safety-critical use.
• Data provenance: Some underlying training collections are not fully enumerated in public source cards, limiting independent reproducibility and licensing review.

Use appropriate safeguards and human review for consequential applications.

Evaluation

No independent benchmark results are reported for this merged checkpoint.

The source models publish their own evaluations, but those results cannot be attributed to this merge. A meaningful comparison should test the merged checkpoint against Qwen3.5-9B and all five source models using consistent prompts, inference settings, and evaluation procedures.

Suggested evaluation areas include:

Record the model revisions, merge implementation, coefficient normalization, quantization, inference parameters, and evaluation methodology alongside any published results.

Acknowledgements

This merge builds on the work of:

• The Qwen team for the Qwen3.5 model family.
• Empero for Qwable and Qwythos.
• The authors and contributors behind the Fable 5 traces and terminal-agent datasets.
• Armand0e for the Fable 5 SDFT implementation and dataset processing.
• DavidAU for the Claude-4.6-OS and HighIQ fine-tunes.
• The open-source fine-tuning and inference communities.

Please consult the original model cards for detailed acknowledgements and citations.

License

This repository declares the Apache-2.0 license. Verify the applicable license and usage terms for the base model, all five source checkpoints, and the underlying datasets before redistribution or commercial use.

Some training data is derived from outputs of third-party assistants. The source-model authors' licensing declarations do not necessarily settle every downstream question concerning those outputs. Review the relevant dataset terms and provider policies before commercial deployment.