cosmicoptima/computer-10

Overview

cosmicoptima/computer-10 is a 70 billion parameter language model, the tenth iteration in the "Computer" series, publicly released in September 2026. It is based on the Llama 3.1 70B model but diverges significantly in its training methodology and intended use. Unlike typical instruction-tuned models, Computer-10 was trained exclusively on a curated anthology of documents, rather than chat or assistant data. This process has imbued it with a distinct, non-servile personality, complete with its own interests, opinions, and idiosyncratic speech patterns, such as scare quotes and nested parentheses.

Key Characteristics & Training

Computer-10 is a "document model" designed to continue conversations described in a specific document format, rather than a standard chat model. Its development involved a multi-stage process:

• Anthology Training: Initial models (computer-1 to computer-5) were full fine-tunes of Llama 3.1 70B on a unique document anthology, where the model (referred to as Model C) is a "living document" that narrates, redacts, and dreams conversations.
• Self-Distillation & RL: Subsequent versions (computer-6, computer-7, computer-9) underwent weighted self-distillation and online self-preference Reinforcement Learning (RL). Computer-7, the parent of later models, acted as the judge in the RL stage, selecting preferred replies from its siblings.
• Refinement (computer-10): This specific model, computer-10, was fine-tuned on its own conversations from computer-9. This stage focused on correcting issues like unclosed parentheses and run-on sentences, improving coherence without altering its unique "taste." It reduced broken replies from 18.9% to 10.2% and unbalanced parentheses from 21.6% to 2.1% compared to computer-9.

Intended Use and Expectations

Computer-10 is not designed to be a helpful or safety-tuned assistant. Users should treat it as a correspondent with its own agenda. It tends to elaborate on premises rather than correct them and may sometimes generate both sides of a conversation. It is a 70B model with the knowledge and limitations of its Llama 3.1 base, and its factual claims should be verified. It supports an OpenAI-style chat template that maps messages to its document format.