Deep Cogito has launched a new lineup of open-source large language models (LLMs) that it claims outperform all similarly sized alternatives, marking a significant step forward in the evolution of artificial general intelligence (AGI).
Next-Gen Open LLMs: Scaling Intelligence with Fewer Resources
Based in San Francisco, Deep Cogito has introduced preview versions of LLMs in 3B, 8B, 14B, 32B, and 70B parameter sizes. According to the company, each model delivers superior performance over current open-source peers such as LLAMA, DeepSeek, and Qwen across widely accepted benchmarks.
In an impressive feat, Deep Cogito’s 70B model outperforms Meta’s recently released Llama 4 109B Mixture-of-Experts (MoE) model on several evaluations.
IDA: Reimagining AI Training with Iterated Distillation and Amplification
At the core of these advancements lies a new training strategy called Iterated Distillation and Amplification (IDA). This technique promotes a self-improving feedback loop where models refine their own intelligence iteratively.
The IDA process consists of two main phases: amplification, which allows the model to generate superior outputs using increased computation, and distillation, where these enhanced capabilities are absorbed back into the model’s architecture. This methodology aims to overcome the limitations of traditional training paradigms that rely heavily on human oversight or larger ‘teacher’ models.
Deep Cogito highlights that IDA enables more scalable and cost-efficient development compared to Reinforcement Learning from Human Feedback (RLHF) and distillation from massive models. For context, the company’s 70B model was developed in just 75 days by a compact team.
Performance Benchmarks: Cogito vs. The Rest
The Cogito models, built on LLAMA and Qwen checkpoints, are optimized for coding, function calling, and agent-like capabilities. A standout feature is their dual-mode functionality—models can respond directly or engage in self-reflection before answering, much like reasoning-based systems.
Tests across standard benchmarks such as MMLU, GSM8K, ARC, and MATH reveal that Deep Cogito’s models consistently outperform rivals of the same size. The Cogito 70B model, for instance, scored 91.73% on MMLU in standard mode, outpacing Llama 3.3 70B by over 6%, and 91.00% in reasoning mode, surpassing DeepSeek R1 Distill 70B by 4.4%.
Benchmark comparisons for the 14B models show similar superiority over Alibaba’s Qwen and DeepSeek R1—a notable achievement as both companies are leading players in the open-source LLM ecosystem. For more on how DeepSeek is reshaping AI alignment, check out this related article: DeepSeek Redefines AI Alignment with Human-Centric Models.
What’s Next for Deep Cogito?
Despite the strong showing, Deep Cogito labels this release as a ‘preview’ and plans to iterate further. The company intends to launch updated checkpoints and larger MoE models in the 109B, 400B, and 671B parameter ranges in the coming months—all of which will remain open-source.
By combining advanced reasoning and iterative self-improvement, the company believes IDA can serve as a foundational strategy for scaling toward general superintelligence. As the AI field continues to evolve rapidly, Deep Cogito’s bold claims and open approach could set a new standard for transparency and performance.
Stay tuned as the race toward AGI accelerates, and new paradigms like IDA challenge the status quo in AI development.
