Empero AI Releases Qwythos-9B-Claude-Mythos-5-1M-GGUF: A 1M Context Reasoning Model with Multimodal Capabilities

What Changed

Empero AI has released the GGUF quantizations for their Qwythos-9B-Claude-Mythos-5-1M model. This release, designated as v3, includes critical hotfixes for the chat template, addressing issues such as preserved reasoning, adaptive thinking, looping during long generation traces, and agentic use in harnesses like OpenCode, Abacus, Hermes, and Claude Code. The GGUF files are designed for compatibility with llama.cpp, Ollama, LM Studio, jan, and KoboldCpp.

The Qwythos-9B model is a full-parameter reasoning model that underwent post-training on over 500 million tokens derived from high-quality Claude Mythos and Claude Fable traces. These traces were generated internally by Empero AI's rethink tool, focusing on chain-of-thought reasoning.

Key features of this GGUF release include native function calling, a 1,048,576-token (1M) context window enabled by YaRN rope-scaling, and multimodal image input capabilities inherited from its Qwen3.5-9B base.

Technical Details

Qwythos-9B-Claude-Mythos-5-1M-GGUF is built upon the Qwen3.5-9B architecture. The model's reasoning capabilities are enhanced by post-training on extensive Claude Mythos/Fable chain-of-thought data. It supports the Qwen3.5 specification for native function calling, emitting <tool_call> blocks for integration into tool-use loops.

The 1M context window is achieved through YaRN rope-scaling, which is enabled by default in the GGUF files. This allows for significantly longer input sequences compared to native context limits.

For local deployment, Empero AI provides several quantization options:

Normal Text Weights (v3 replacements):

• Qwythos-9B-Claude-Mythos-5-1M-Q4_K_M.gguf (5.24 GiB): Recommended default for balanced quality and size.
• Qwythos-9B-Claude-Mythos-5-1M-Q5_K_M.gguf (6.02 GiB): Balanced quality/size.
• Qwythos-9B-Claude-Mythos-5-1M-Q6_K.gguf (6.85 GiB): High quality.
• Qwythos-9B-Claude-Mythos-5-1M-Q8_0.gguf (8.87 GiB): Near-lossless.
• Qwythos-9B-Claude-Mythos-5-1M-BF16.gguf (16.69 GiB): Full precision conversion base.

MTP-enabled Text Weights (v3 variants): These include the restored Qwen3.5-compatible MTP head for use with llama.cpp builds supporting MTP draft speculation (e.g., --spec-type draft-mtp).

• Qwythos-9B-Claude-Mythos-5-1M-MTP-Q4_K_M.gguf (5.48 GiB): Recommended MTP default.
• Qwythos-9B-Claude-Mythos-5-1M-MTP-Q5_K_M.gguf (6.26 GiB): MTP, balanced quality/size.
• Qwythos-9B-Claude-Mythos-5-1M-MTP-Q6_K.gguf (7.09 GiB): MTP, high quality.
• Qwythos-9B-Claude-Mythos-5-1M-MTP-Q8_0.gguf (9.11 GiB): MTP, near-lossless.
• Qwythos-9B-Claude-Mythos-5-1M-MTP-BF16.gguf (17.14 GiB): MTP, full precision conversion base.

For image input, a separate vision projector file, mmproj-Qwythos-9B-Claude-Mythos-5-1M-F16.gguf (0.86 GiB), is required. This vision tower is inherited directly from the Qwen3.5-9B base model and was frozen during the SFT process, meaning its vision behavior is identical to the base Qwen3.5-9B's multimodal capabilities. It is interchangeable with any community-built Qwen3.5-9B mmproj-*.gguf.

Recommended sampling parameters for Qwythos, a reasoning model that starts responses with a <think>...</think> block, include temperature 0.6, top_p 0.95, top_k 20, and repeat_penalty 1.05. A max_new_tokens of 16384 is suggested to accommodate the reasoning block and final answer. Greedy decoding and very-low-temperature sampling (T ≤ 0.3) are advised against to prevent repetition loops.

Benchmark Analysis

Empero AI reports that Qwythos-9B demonstrates significant performance improvements over the base Qwen3.5-9B model under matched lm-eval-harness evaluation:

• +34 points on MMLU (Massive Multitask Language Understanding)
• +30 points on gsm8k-strict (Grade School Math 8K, strict matching)
• +19 points on gsm8k-flex (Grade School Math 8K, flexible matching)

In a 7-prompt tool-use harness involving a Python executor and DuckDuckGo search, Qwythos achieved source-cited correct answers on all 7 prompts, including 4 closed-book failure modes from the original review.

Developer Implications

Developers can leverage Qwythos-9B-Claude-Mythos-5-1M-GGUF for applications requiring advanced reasoning, long-context understanding, and multimodal input, particularly in environments supporting GGUF runtimes. The model's native function calling simplifies integration into agentic workflows and tool-use scenarios.

The availability of various quantizations allows for flexibility in deployment, balancing model quality with computational resources. The Q4_K_M quantization is recommended as a starting point for its balance of size and quality. For those utilizing llama.cpp with MTP draft speculation, specific MTP-enabled GGUFs are provided.

For multimodal applications, the mmproj file enables image input, supporting detailed image description, OCR, chart/table reading, UI/document understanding, and basic spatial reasoning. Developers should note that the vision tower was not fine-tuned with image-paired data, so image-grounded reasoning performance mirrors the base Qwen3.5-9B and should be validated for primary vision-driven use cases.

The model is uncensored, which implies a need for application-level safety layers in user-facing deployments where content moderation is necessary. Its reasoning-first approach, indicated by the <think>...</think> blocks, requires developers to account for this structure in prompt engineering and output parsing.

Bottom Line

Empero AI's Qwythos-9B-Claude-Mythos-5-1M-GGUF offers a powerful, locally deployable reasoning model with a 1M token context window and multimodal capabilities. Its strong performance against the Qwen3.5-9B base model in reasoning benchmarks, coupled with native function calling, positions it as a robust option for developers building advanced AI applications. The availability of GGUF quantizations ensures broad compatibility with popular local inference engines, making high-performance AI more accessible for development and experimentation.

Pneumetron