Anubis 70B v1

Overview

Anubis 70B v1 is a large-scale generative language model developed by BeaverAI and presented by TheDrummer. Built as a fine-tuned version of Llama 3.3, the model incorporates 70.6 billion parameters and is positioned as a versatile solution for creative writing, role-playing, and dialogue generation. With its focus on producing natural language interactions, Anubis 70B v1 aims to combine creativity with coherency, offering users engaging and responsive conversational experiences.

Model Architecture and Training

Anubis 70B v1 is built upon the Llama 3.3 foundation, a model architecture recognized for its broad applicability in generative AI tasks. The model contains approximately 70.6 billion parameters, leveraging the BF16 tensor type to optimize numerical precision and computational efficiency during inference. As a fine-tuned variant, Anubis 70B v1 extends the baseline capabilities of Llama 3.3, with a focus on delivering enhanced performance in dialogue and storytelling scenarios. While specific training datasets and techniques for this iteration have not been publicly detailed, the model inherits the robust pretraining framework of its base model, supplemented with additional fine-tuning tailored to creative and narrative-rich outputs, as described in its Hugging Face documentation.

Capabilities and Performance

Users have observed that Anubis 70B v1 is distinguished by a balanced performance profile among Llama 3.3-based derivatives, achieving a synthesis between creativity and coherence. The model excels at generating vivid and contextually appropriate scenarios, producing natural-sounding dialogue and immersive prose. These attributes make it particularly effective for applications where lifelike character interactions are paramount.

A notable strength of Anubis 70B v1 is its aptitude for role-playing tasks and creative fiction, enabling text outputs that emulate the expressiveness and spontaneity of human authorship. Reports from practitioners indicate that the model can occasionally perform on par with or exceed expectations set by its size class, with some comparing its language generation abilities favorably to models of larger scale. Anubis 70B v1 also handles a variety of prompting styles, thriving on clear and detailed instructions, and is compatible with conventionally structured chat templates supporting system, user, and assistant roles as outlined in the Llama 3 Chat template documentation.

Use Cases and Applications

The primary domains for Anubis 70B v1 are creative writing and character-driven interactions, reflecting its tuning towards sustained, engaging dialogue and role-play formats. The model particularly thrives in settings where narrative depth and character voice are central, such as interactive fiction, simulation games, and online collaborative storytelling. A commonly preferred structure utilizes a blend of actions, dialogue, thoughts, and narration, often presented in the first-person perspective to maximize immersion and relatability for users.

Additionally, Anubis 70B v1 has seen use in conjunction with custom character cards, which provide defined personas for interaction and further enhance the quality of the generated exchanges. Guidance for optimal use highlights the importance of well-structured prompts and leveraging the chat role templates to achieve coherent, high-fidelity responses.

Deployment, Compatibility, and Technical Details

For integration and deployment in various workflows, Anubis 70B v1 is available in several quantized formats, allowing users to balance resource requirements with performance. The recommended iMatrix GGUF version is maintained here, as the original GGUF link is no longer active. This flexibility facilitates experimentation and accommodates diverse technical environments. Support for the Safetensors format underlines the model’s compatibility with modern storage and security best practices.

The model employs the Llama 3.3 Chat prompt structure, making it straightforward to integrate within standard conversational AI frameworks. While no official hardware or inference provider requirements have been specified, the model is intended for research and creative purposes, and further technical details can be explored on its Hugging Face hub entry.

Community, Licensing, and Ecosystem

Anubis 70B v1 is developed and maintained in a spirit of open collaboration, with community engagement occurring primarily through platforms such as the BeaverAI Discord server. While the specific license governing the use and distribution of Anubis 70B v1 has not been explicitly published, interested users are encouraged to consult the model’s official documentation for the latest information and guidance.

Within its ecosystem, Anubis 70B v1 is often compared to other fine-tuned Llama models, including larger alternatives like Behemoth, though it distinguishes itself by offering a strong balance of efficiency and capability suitable for creative and conversational tasks. Additional support resources, including example character cards such as Audrey on CharacterHub, are widely adopted by the community to further personalize and customize interaction experiences.

References and Helpful Links

• Anubis 70B v1 on Hugging Face
• Recommended GGUF Quantization (iMatrix)
• BeaverAI Community Discord
• Safetensors documentation
• Llama 3 Chat Template Information
• Audrey Character Card Example

For further exploration, researchers and users can consult the original model documentation and resources for updates, technical notes, and community discussions.