【AIbase Report】Recently, a little-known but highly anticipated San Francisco startup, Deep Cogito, released its Cogito v2 series of large language models (LLMs), aiming to break through in the crowded open-source AI market. Unlike traditional parameter stacking strategies, this company, founded by former Google engineers, bets on "machine intuition" and self-improving reasoning capabilities, creating an AI model that can "learn while using."
The model not only answers questions, but also learns "how to answer questions"
The Cogito v2 series includes four models with parameter sizes ranging from 70B to 671B, divided into dense models (Dense) and mixture-of-experts models (MoE), both of which are now available on platforms such as Hugging Face and Together AI. The flagship model, Cogito v2-671B MoE, is called "the most efficient thinker," with a reasoning path 60% shorter than DeepSeek R1, while its performance can match or even exceed Qwen1.5-72B and Claude4Opus.
Its core technology lies in the fact that the model performs "introspective reasoning" during operation and distills these reasoning paths back into the model weights, forming internalized intuition. This mechanism allows the model to become "smarter" with each reasoning process, similar to how AlphaGo strengthens its strategy through games.
Practical Testing of Reasoning Power: Faster and Shorter Paths
Deep Cogito has released multiple test cases to verify its "machine intuition":
In mathematical problems, Cogito671B accurately reaches conclusions with a reasoning chain as short as 100 tokens, while DeepSeek R1 used over 200 tokens.
In legal reasoning, it uses a two-step logical structure to produce clear conclusions, exceeding many models and even the performance of real law master's students.
In the classic family logic question "Is Alice Charlie's grandmother?" Cogito v2 successfully avoided the pronoun confusion trap and accurately output "grandmother."
A More Cost-effective Training Approach, Challenging the Myth of Millions in Budget
Although the Cogito v2 model scale is huge, Deep Cogito claims that the total training cost for eight models is less than $3.5 million, which contrasts sharply with the research and development costs of OpenAI and Anthropic, which often reach hundreds of millions of dollars.
CEO Drishan Arora said, "Better models are not about training more data, but training more meaningful data." This is precisely the key to Cogito's breakthrough in reasoning tasks.
Continuation of the Open Source Philosophy, Building a "Evolutionary Model System"
The Cogito v2 models are currently available for download or API calls via platforms such as Hugging Face, Baseten, RunPod, and Unsloth. To support lightweight deployment scenarios, Cogito671B also offers an FP8 quantized version, enabling large models to run with lower hardware requirements, improving inference efficiency with only a slight decrease in accuracy.
More importantly, Deep Cogito promises that all models will be open source and will continue to iterate and optimize, forming a new model training path centered on "reasoning chain feedback + self-improvement."
Currently, Cogito v2 has received attention and support from well-known institutions such as Benchmark and South Park Commons, and is seen as a dark horse in the open-source AI field.