Cogito AI: Free & Revolutionary
- SAN FRANCISCO, April 13, 2025 – Deep Cogito, a new artificial intelligence research startup based in San Francisco, has emerged from stealth mode with teh release of its...
- The initial Cogito V1 release includes five models of varying sizes, ranging from 3 billion to 70 billion parameters.
- The Cogito models are built upon the Llama 3.2 architecture from Meta, but Deep Cogito has implemented its own training approach called IDA, or iterated distillation and amplification.
Deep Cogito Launches Open-Source AI Language Models, Challenges Competitors
Table of Contents
SAN FRANCISCO, April 13, 2025 – Deep Cogito, a new artificial intelligence research startup based in San Francisco, has emerged from stealth mode with teh release of its first series of open-source language models, dubbed Cogito V1.
The initial Cogito V1 release includes five models of varying sizes, ranging from 3 billion to 70 billion parameters. The company reports these models were developed by a small team in just 75 days.
The Cogito models are built upon the Llama 3.2 architecture from Meta, but Deep Cogito has implemented its own training approach called IDA, or iterated distillation and amplification.
IDA: A Novel Approach to AI Training
Traditional AI model training frequently enough relies on human feedback, known as Reinforcement Learning from Human Feedback (RLHF), or uses the output of a more powerful “teacher” model to guide learning.
Deep Cogito’s IDA method takes a diffrent approach. Instead of relying on external feedback, the model generates multiple response options, selects the best one, and then uses that selection to refine its reasoning. This iterative process allows the model to improve without direct human intervention.
Drishana Arory,a former Google engineer,has described IDA as a “linguistic AlphaGo,” drawing a parallel to the AI model that mastered the game of Go through self-play.
Performance Benchmarks Show Promise
Deep Cogito has released benchmark data indicating that its models outperform comparable open-source alternatives in several areas.
Cogito 3B benchmark”>
The benchmarks suggest that the Cogito models perform notably well in “reasoning” mode, where they are given more time to process and generate responses. However, they lag behind other models, such as Deepseek, in mathematics-related tasks.
Future Plans: Larger Models and Continued Open-Source Commitment
Deep Cogito plans to expand the Cogito family with even larger models, including versions with 109 billion, 400 billion, and 671 billion parameters. The company has pledged to keep all its models open-source, a move that could attract developers and organizations seeking powerful AI solutions without the constraints of closed systems.
The models are available for trial use.
Deep Cogito AI: Your Burning Questions Answered
Welcome! Let’s dive into Deep Cogito, the new AI player making waves in the tech world. I’ll break down everything you need to know, answering your likely questions in a clear, concise, and helpful way.
Q: what is Deep Cogito?
A: Deep Cogito is a newly emerged AI research startup based in San Francisco. They’ve just launched thier first series of open-source language models,aptly named “Cogito V1.”
Q: What models are included in the Cogito V1 release?
A: The initial release features five models, each with a different number of parameters, ranging from 3 billion to 70 billion parameters. They were developed in just 75 days by a smaller team!
Q: What architecture are Cogito models based on?
A: The Cogito models leverage the Llama 3.2 architecture from meta. However, Deep Cogito has implemented its unique training approach.
Q: What’s unique about Deep Cogito’s training approach?
A: Deep Cogito uses a novel method called IDA,or Iterated distillation and Amplification.
* IDA Explained: Instead of relying on human feedback (RLHF) or a ”teacher” model, IDA allows the model to generate multiple response options. It then selects the best one and uses it to refine its reasoning in an iterative process. This self-improvement approach is a key differentiator.
Q: How does IDA compare to othre AI training methods?
A: Traditional approaches often use human feedback to guide the AI’s learning. IDA differs by enabling the model to improve without direct human input. Drishana Arory, a former Google engineer, even described it as a “linguistic AlphaGo,” drawing a parallel to the AI model that mastered the game of Go through self-play.
Q: how do the Cogito models perform?
A: Benchmarks released by deep Cogito suggest that cogito models outperform comparable open-source alternatives in several areas, especially in “reasoning” mode.
Q: Can you summarize the benchmark results?
A: I sure can. Here’s a quick comparison of what we know:
| Model | Performance (Reasoning) | Performance (Mathematics) |
|---|---|---|
| Cogito models | Notably well | Lag behind other models |
| Comparable Open-Source Alternatives | Outperformed by Cogito in several areas | Not mentioned |
| Deepseek (mentioned as comparison) | Not mentioned | Performance surpasses Cogito |
Q: What are Deep Cogito’s future plans?
A: Deep Cogito plans to expand the Cogito family with even larger models, including versions with 109 billion, 400 billion, and 671 billion parameters.They’ve also committed to keeping all their models open-source.
Q: Why is open-sourcing vital?
A: Keeping the models open-source is a strategic move that coudl attract developers and organizations. This approach allows them to harness powerful AI solutions without restrictions.
Q: Where can I try out the Cogito models?
A: The article states that the models are available for trial use.