cactopus/Omega_Sapphira_Joyous-L3.3-70B-v1.0

TEXT GENERATIONPricing:Input $2.88 / Output $2.88Concurrent Unit Cost:4Model Size:70BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 3, 2026License:llama3.3Architecture:Transformer Featherless Exclusive Cold

Omega-Sapphira-Joyous-L3.3-70B-v1.0 is a 70 billion parameter experimental merge of three Llama 3.3 fine-tunes, developed by cactopus, featuring a unique depth-graded blending method. This model's ancestry shifts across its 80 layers, with varying contributions from Omega, Sapphira, and Joyous models to different sub-blocks. It is designed for creative writing, offering a warmer prose register, and is unaligned for adult fiction with a 32768 token context length.

Loading preview...

Model Overview

Omega-Sapphira-Joyous-L3.3-70B-v1.0 is an experimental 70 billion parameter language model developed by cactopus, built upon the Llama 3.3 architecture. This model distinguishes itself through a novel depth-graded merging technique, where the contributions of its three parent models (Omega, Sapphira, and Joyous) vary across the 80 layers and between attention and feed-forward blocks. This results in a dynamic ancestry profile throughout the network, rather than a fixed ratio merge.

Key Characteristics

  • Dynamic Ancestry: The model's composition changes layer by layer, with Omega contributing significantly to structural coherence (self_attn) and Sapphira to prose style (MLP).
  • Unaligned Nature: Inherits the unaligned characteristics of its parent models, suitable for generating explicit and violent material without refusal.
  • Context Length: Supports a context length of 32768 tokens, with practical usage up to 40448 tokens on specific hardware configurations.
  • Tokenizer: Uses the tokenizer from allura-org/Llama-3.3-70B-Joyous, which includes add_bos_token, pad_token_id, and generation_config.json.

Intended Use Cases

  • Creative Writing: Optimized for generating creative text with a "warmer prose register," trading some structural discipline for expressive style.
  • Adult Fiction: Designed to engage with explicit and violent content, making it suitable for mature storytelling applications.

Important Considerations

This model is superseded by Omega-Sapphira-Joyous-L3.3-70B-v1.1, which addresses an over-representation of the Joyous component in this version. For a more stable and well-behaved merge, users are directed to Omega-Sapphira-L3.3-70B-v1.3. Users should be aware of the model's unaligned nature and are responsible for the content they generate.

Finetunes
2 models