Empero AI Releases Qwythos-9B-Claude-Mythos-5-1M Open-Weight Model

Qwythos-9B-Claude-Mythos-5-1M is a 9 billion parameter full-parameter reasoning model fine-tuned from a deeply uncensored Qwen3.5-9B base and post-trained on Claude Mythos and Fable traces.

Empero AI announced the release of its Qwythos-9B-Claude-Mythos-5-1M model on June 19, 2026, positioning it as the company's flagship open-weights offering. The model supports image-text-to-text modalities though the vision tower remains frozen during fine-tuning, allowing focus on text-based reasoning enhancements. Released under the Apache-2.0 license, it includes GGUF quantizations to facilitate deployment across various hardware setups. This development comes at a time when the AI community seeks more accessible alternatives to closed frontier systems, particularly in areas requiring extended context and reliable tool interaction. The base model originates from Qwen3.5-9B, selected for its uncensored nature that permits broader post-training applications without inherent restrictions. Empero AI focused on full-parameter updates rather than adapter-based methods to maximize the integration of new reasoning patterns throughout the entire network.

What background and context surround the Qwythos-9B-Claude-Mythos-5-1M release?

The development of Qwythos-9B-Claude-Mythos-5-1M stems from Empero AI's efforts to create a reasoning model that can handle complex tasks through self-correction mechanisms. The company utilized its internal rethink tool to generate chain-of-thought processes from session logs of Claude Mythos and Claude Fable. This post-training on more than 500 million tokens aims to instill advanced reasoning patterns into the smaller 9B parameter model. Such an approach allows the model to achieve performance levels that rival larger systems in specific domains. The choice of an uncensored base facilitates the integration of these synthetic traces without conflicts that might arise from more restricted models. Community discussions on platforms like Reddit have highlighted the potential for self-hosted applications given the model's open nature and the availability of quantized versions for consumer hardware.

Prior models in the open-weight space often struggled with consistent tool use over long contexts or required extensive prompt engineering to maintain coherence. Empero AI addressed these gaps by distilling behaviors from higher-performing closed systems into a compact form factor. The post-training process emphasized quality over quantity in the synthetic dataset, drawing directly from high-fidelity traces that include detailed reasoning steps. This method contrasts with traditional fine-tuning that relies solely on human-annotated data. The resulting model demonstrates improved coherence when handling multi-step problems that involve both reasoning and external function calls. Release notes from Empero indicate that the internal tool played a central role in scaling the generation of these traces efficiently.

What new capabilities does the model introduce in detail?

Qwythos-9B-Claude-Mythos-5-1M introduces native function calling aligned with the Qwen3.5 specification, enabling direct integration with external tools without custom adapters. The model achieved 7 of 7 successes on a tool-use harness covering math, cybersecurity, and clinical pharmacology queries, incorporating self-correction loops that allow it to revise outputs based on intermediate results. This capability extends to scenarios where initial tool responses require follow-up calls or adjustments. The 1,048,576 token context window, activated by default through YaRN rope-scaling, supports processing of extensive documents or extended dialogue histories in a single forward pass. Image inputs are supported at the modality level, though the vision components stayed frozen to prioritize text reasoning gains.

The self-correction feature manifests during tool interactions by allowing the model to detect inconsistencies in its own outputs and initiate corrective actions. This reduces the need for external oversight in automated workflows. Performance metrics show substantial lifts on standard benchmarks, with the model demonstrating reliable behavior across the tested domains. The GGUF quantizations further lower the barrier for local inference, making the capabilities available to users without access to large GPU clusters. These features collectively position the model as a practical option for agentic applications where long context and tool reliability are essential.

What are the technical specifics of the implementation?

The architecture retains the core transformer structure of the Qwen3.5-9B base while applying full-parameter updates during the post-training phase. YaRN rope-scaling modifies the positional embeddings to support the extended context length without requiring changes to the underlying attention mechanism. The post-training dataset consisted of over 500 million tokens derived from Claude Mythos and Claude Fable session logs, with chain-of-thought sequences produced by Empero AI's rethink tool. This synthetic data emphasizes step-by-step reasoning that the model internalizes for improved performance on downstream tasks. Function calling follows the exact format specified in the Qwen3.5 documentation, ensuring compatibility with existing tool-calling frameworks.

Benchmark performance comparison between base and fine-tuned model

Metric	Base Qwen3.5-9B	Qwythos-9B-Claude-Mythos-5-1M
MMLU	0.232	0.575
GSM8K strict exact match	0.510	0.810

The training pipeline began with selection of the uncensored Qwen3.5-9B checkpoint to maximize flexibility in absorbing the new data distribution. Subsequent stages involved supervised fine-tuning on the synthetic traces followed by alignment steps to reinforce tool-use behaviors. The decision to freeze the vision tower during fine-tuning preserved the original multimodal capabilities while directing computational resources toward text and reasoning improvements. GGUF quantization formats were prepared post-training to support efficient inference on consumer devices. These technical choices reflect a balance between capability gains and practical deployability.

1. Selection of the deeply uncensored Qwen3.5-9B base model as the foundation for full-parameter updates.
2. Generation of over 500 million tokens of high-quality Claude Mythos and Claude Fable traces from session logs.
3. Application of the internal rethink tool to produce detailed chain-of-thought sequences for each trace.
4. Implementation of YaRN rope-scaling to activate the 1,048,576 token context window by default.
5. Fine-tuning for native function calling compatibility with the Qwen3.5 specification.
6. Release of GGUF quantizations alongside the base weights under the Apache-2.0 license.

What market and stakeholder implications arise from this release?

The availability of a 9B model with 1M token context and verified tool-use performance under an open license expands options for developers building agentic systems without reliance on proprietary APIs. Organizations focused on data privacy can deploy the model locally while benefiting from extended context for document analysis or multi-turn interactions. The performance improvements on MMLU and GSM8K suggest that distillation from stronger models can narrow the gap between small and large parameter counts in targeted areas. Stakeholders in the open-source community gain a new reference point for evaluating post-training techniques that combine synthetic data with self-correction mechanisms.

Hardware vendors may see increased demand for systems capable of running 9B models at long contexts, particularly with the inclusion of quantized variants. Academic researchers can study the effects of large-scale synthetic trace distillation on smaller architectures. The Apache-2.0 terms permit commercial use and modification, potentially accelerating integration into enterprise tools. Overall, the release contributes to a trend of making advanced capabilities more distributed rather than concentrated in closed models.

How have expert reactions manifested following the announcement?

We just shipped Qwythos-9B-Claude-Mythos-5-1M — our biggest open-weights model to date, and the new flagship over on Hugging Face. It's a full-parameter reasoning model built on a deeply uncensored Qwen3.5-9B base, post-trained on north of 500 million tokens of Claude Mythos and Claude Fable traces, with the chain-of-thought generated in-house by our rethink tool.kodee, Empero Research

Reactions in technical forums have centered on the practical utility of the tool-calling harness results and the context extension. Users testing the GGUF versions report stable performance during extended sessions. The combination of self-correction with native function calling has drawn attention as a step toward more autonomous agent behaviors in open models. Empero AI's emphasis on the internal rethink tool for trace generation has prompted discussions about scalable methods for creating high-quality synthetic reasoning data.

What developments can be expected next in this area?

Future iterations from Empero AI may incorporate additional modalities or further scale the post-training dataset while maintaining the compact parameter count. The success of the current approach could encourage similar distillation pipelines across other base models in the open ecosystem. Continued refinement of self-correction techniques may lead to higher reliability in complex multi-tool workflows. The 1M context capability opens avenues for applications in long-form content generation and analysis that were previously limited to larger closed models. Stakeholders should monitor updates to the Hugging Face repository for new quantizations or fine-tunes that build on the released weights.