New open source AI company Deep Cogito releases first models and they’re already topping the charts

The launch of the new AI research in San Francisco has been officially revealed with COGITO today V1, open source large language models (LLS), meta’nin llama 3.2 is equipped with hybrid justification capabilities.

The company aims to go beyond the current human controller restrictions by denying the Borders of the EI’s borders as an actorial and internal regulatory strategies. As a result, AI is smarter than all people in all domains – but the company says “All the models we create will be open.”

Director General of the Deep Cogito and co-founder Drizan Arora – a large program engineer in Google, a former Software Engineer who said that he led the Google Geni Model (LLM) modeling –He also said in an article in X They are the most powerful open models on their scale, including “Llama, DeepSeek and Qwen.”

The initial model line includes five bases: 3 billion, 8 billion, 14 billion, 32 billion, 32 billion and 70 billion parameter, now available in the AI ​​Code Society Society Hug face, Activation and application programming interfaces (API) Fireworks and Together ai.

Are available under Lülama Licensing Terms Third-party enterprises for the use of trade – can launch them in paid products – to 700 million monthly users, in this point you need to get a paid license from Meta.

The company plans to release further up to 671 billion parameters – in the coming months.

Arora describes the company’s training approach, iterated distillation and strengthening and strengthening and strengthening (IDA), human opinion (RLHF) or teacher model distillation (RLHF) as a novel alternative (IDA).

The main idea behind the Ida, to allocate more calculation for a model to create improved solutions, then take the improved justification process to your settings – Create a feedback loop for the ability to increase. Arora is similar to this approach to the natural language, Google Alphago’s self-playing strategy.

COGITO models are an open source and fireworks are open sources and open sources to download through AI and AI-submitted by AI or to download through APIs. Each model supports standard mode for direct answers and reasoning regimens that reflect the internal before answering the units.

Evaluation and evaluation

The company shared COGito models in general knowledge, mathematical thinking and extensive assessment results than in the source peers. Highlights include:

• COGITO 3B (default) over Llama 3.2 3b 6.7 percent points in the MML (65.4% – 58.7%) and 18.8 points in Hellaswag (81.1% and 62.3%).
• In Reason mode, Cogito 3B 72.6% in the MMLI, 84.2% in the ARC, exceeding its standard mode performance and the effect of reflecting Ida.
• COGITO 8B (default) 80.5% of Puan in MMLU is superior Call 3.1 8B 12.8 points. He also leads to more than 11 points in MMLU-PRO and reaches 88.7% in the arc.
• In Reason mode, COGITO 8B 83.1% in MMLU, reaches 92.0% in the arc. Is superior DeepSeek R1 Dusty 8B COGITO is in almost every category in another category of significantly low (60.2% and 80.6%).
• COGITO 14B and 32B Models Outperform Qwen2.5 In total of 2-3 percent of the colleagues in total criteria COGITO 32B (justification) It reaches 90.2% in MLZ and reaches 91.8% to the mathematics criterion.
• COGITO 70B (default) over Call 3.3 70B 6.4 points in MMLU (91.7% – 85.3%) and exceeds Llama 4 Scout 109B In the cumulative benchmark scores (54.5% 53.3%).
• Against DeepSeek R1 Distily 70B, COGITO 70B (justification) Posts are stronger results, MML and MML and MGSM 91.0%, 91.0% and 92.7%.

COGITO models generally show the highest performance to reveal basic performances, despite some trading-offs emerge – especially in mathematics.

For example, COGITO 70B (standard) mathematics and GSM8K, COGITO 70B (justification) roads in math in math in math (89.3% and 89.0%)

Interior caller

In addition to the general criteria, Deep Cogito, local tool evaluated the models of challenge performance – an increasing priority for agents and API-inteated systems.

• COGITO 3B Local tool supports calling tasks locally (simple, parallel, multiple and parallel-multiple), whereas Llama 3.2 3b Does not support the instrument call.
• COGITO 3B is more than 91% on 92.8% in simple vehicle calls and more than 91%.
• COGITO 8B, all tools are more than 89% along call types, significantly superior Call 3.1 8BVaries between 35% and 54%.

These improvements are not only modeling architectural and learning information, but also the position of the position of the positions that there are no many initial models.

Looking forward

Deep COGITO, 109B, 400B and 671B plans to release further scale models, including mixing-expert options in the parameters of 400B and 671B. The company will also continue to update existing training points with an expanded training.

The company places the Ida methodology as a long-term way to eliminate dependence on human or static teacher models and expanding itself.

Arora highlights

Deep Cogito’s research and infrastructure partnership, FACE, Runpod, Fireworks AI include teams of embracing AI and activ. All released models are open sources and are now available.