This website uses cookies to anonymously analyze website traffic using Google Analytics.
AcceptDecline

Models / Cogito

Cogito

Intelligence that improves itself. Reasoning made efficient.

Cogito v2 models achieve frontier performance through iterative self-improvement, delivering 60% shorter reasoning chains than competing models at a fraction of the cost. Build superintelligence with open-source IDA.

Try Cogito Models
Contact Sales

Get Started in Minutes

Drop-in OpenAI replacement—no code changes, no surprises on your bill. Switch from closed models to Cogito instantly with OpenAI-compatible endpoints on Together AI.

# Install the Together AI library
pip install together

# Get started with Cogito v2
from together import Together
client = Together()

response = client.chat.completions.create(
    model="deepcogito/cogito-v2-preview-llama-70B",
    messages=[
        {
            "role": "user",
            "content": "Explain quantum computing with reasoning"
        }
    ],
    # Enable reasoning mode
    enable_thinking=True
)
print(response.choices[0].message.content)
View API Docs

Why Cogito for Superintelligence?

The first models to achieve true iterative self-improvement

‍

Built on Iterated Distillation and Amplification (IDA), Cogito models don't just search longer—they develop better intuition through self-improvement, paving the way to superintelligence.

  • ITERATIVE SELF-IMPROVEMENT

    • First reasoning models to improve core intelligence, not just search time
    • Models develop stronger "intuition" through distillation of reasoning processes
    • 60% shorter reasoning chains than DeepSeek R1 with superior performance

    BREAKTHROUGH EFFICIENCY

    • Complete model family trained for <$3.5M total cost
    • Significantly more efficient than capital-intensive approaches
    • Proves superintelligence research is accessible, not just for big tech

    OPEN SUPERINTELLIGENCE

    • All models released under open license for commercial use
    • Complete transparency in reasoning process with <think> tags
    • Build on our research or deploy anywhere without restrictions

Meet the Cogito Model Family

From efficient reasoning to frontier intelligence, choose the Cogito model that fits your superintelligence journey.

Cogito 70B

Efficient Hybrid Reasoning

Starting at $0.88/1M tokens

  • 70B

    Dense

  • 128K

    Context

  • 30+

    Languages

  • Hybrid

    Mode

Key Strengths:

  • Superior intuition vs Llama 3.1-70B

  • Strong reasoning capabilities

  • Cost-effective deployment

Try it out

Cogito 109B MoE

Advanced MoE Reasoning

Starting at $0.59/1M tokens

  • 109B

    MoE

  • 10M

    Context

  • Transfer Learning

    Vision

  • Hybrid

    Mode

Key Strengths:

  • Emergent visual reasoning

  • Extended context

  • Efficient MoE architecture

Try it out

Cogito 405B

Frontier Intelligence

Starting at $3.50/1M tokens

  • 405B

    Dense

  • 128K

    Context

  • Frontier

    Performance

  • Advanced

    Reasoning

Key Strengths:

  • Approaches closed frontier models

  • Superior non-reasoning mode

  • Advanced distillation

Try it out

Cogito 671B MoE

Superintelligence Preview

Starting at $1.25/1M tokens

  • 671B

    MoE

  • Matches

    DeepSeek v3

  • 60%

    Shorter Chains

  • Both

    Modes Improved

Key Strengths:

  • Strongest open model globally

  • Improved reasoning intuition

  • Proof of concept for superintelligence

Try it out

Breakthrough Technical Innovations

Cogito models introduce revolutionary approaches to building superintelligence through iterative self-improvement.

  • Iterated Distillation & Amplification (IDA)

    Scalable training recipe where models improve their own intelligence prior through distillation of reasoning discoveries back into parameters. Unlike pure search scaling, this develops true "intuition."

    First successful implementation for general reasoning

  • Iterative Policy Improvement

    Models don't just search longer—they learn from each search to start closer to solutions. Distills reasoning trajectories to build stronger priors for future reasoning.

    60% shorter reasoning chains than competitors

  • Hybrid Architecture

    Single models that can operate in standard LLM mode or reasoning mode. Seamless switching between fast inference and deep reasoning without separate model deployment.

    Best of both worlds in one model

  • Signal for Thinking Process

    Unlike methods that only provide signal on final outputs, we provide training signal for the thinking process itself during distillation, discouraging meandering.

    Efficient reasoning without wandering

  • Transfer Learning Vision

    Models learn visual reasoning capabilities through pure transfer learning from text training, with no explicit multimodal examples in training pipeline.

    Emergent cross-modal reasoning abilities

Real Performance Benchmarks

See how Cogito models stack up against competition on verified benchmarks, achieving frontier performance through better intuition, not just longer search.

Qwen3-235B-A22B

Cogito 671B MoE vs Competition (Non-Reasoning):

Model

MMLU

MMLU-Pro

GPQA Diamond

GSM8K

MATH

Cogito 671B MoE

87.60

80.53

69.70

96.36

92.40

DeepSeek-v3

86

77.13

55.05

96.24

91.80

Claude 4 Opus

92

85.73

70.71

97.19

94.79

GPT-4.1

89.60

82.13

72.22

96.13

92.10

Cogito 671B MoE vs Competition (Reasoning)

Model

MMLU

AIME 2024

GPQA Diamond

GSM8K

MATH

Cogito 671B MoE

91.33

96.21

76.02

96.21

98.17

DeepSeek-R1

90.20

95.75

64.65

95.75

95.59

o3

93.60

97.19

81.03

97.19

98.10

Claude 4 Opus

92.93

97.12

75.76

97.12

97.60

Deploy on Together AI

Access Cogito models through Together's optimized inference platform with enterprise-grade security and performance guarantees.

  • Serverless Endpoints

    Pay-per-token pricing with automatic scaling. Perfect for getting started or variable workloads.

    Best for:

    • Prototyping and development

    • Variable or unpredictable traffic

    • Cost optimization for low volume

    • Getting started quickly

    Cogito 70B:
    $0.88/1M tokens

    ‍    Cogito 109B MoE::
    $0.18 input/1M tokens
    $0.59 output/1M tokens

    Cogito 405B:
    $3.50/1M tokens

    Cogito 671B MoE:
    $1.25/1M tokens

    Try Serverless

  • On-Demand Dedicated

    Dedicated GPU capacity with guaranteed performance. No rate limits. Built for production.

    Best for:

    • Production applications

    • Extended model library access

    • Predictable latency requirements

    • Enterprise SLA needs

    Cogito 405B:
    $0.67/minute (8x H200)

    Cogito 671B MoE:
    ‍$0.67/minute (8x H200)

    Deploy Endpoint

  • Monthly Reserved

    Committed GPU capacity, enterprise features and volume discounts. Optimized for scale.

    Best for:

    • High-volume committed usage

    • Enterprise security requirements

    • Priority hardware access

    • Maximum cost efficiency

    Reserved GPU pricing:
    Starting $0.98/hr

    Volume Discounts:

    Up to 40% savings

    Contact Sales

Enterprise-Grade Security

Your data and models remain fully under your control with industry-leading security standards.

  • SOC 2 Type II


    Comprehensive security controls audited by third parties.

  • HIPAA Compliant

    Healthcare-grade data protection for sensitive workloads.

  • Model Ownership

    You own your fine-tuned models and can deploy anywhere.

  • US-Based Infrastructure

    Models hosted on secure North American servers with strict data sovereignty controls.

Frequently Asked Questions

How does Cogito's self-improvement differ from longer reasoning?

Cogito models don't just search longer—they improve their core intelligence through Iterated Distillation and Amplification. Each reasoning discovery is distilled back into model parameters, making future reasoning more efficient. This results in 60% shorter reasoning chains while achieving better performance.

What is Iterated Distillation and Amplification (IDA)?

IDA is our scalable training approach for superintelligence. Models generate improved reasoning through search, then distill those discoveries back into their parameters. This creates a feedback loop where models become inherently more intelligent, not just better at searching.

Can I use Cogito models commercially?

Yes! All Cogito models are released under permissive open source licenses (MIT, Llama) that allow commercial use. You own your fine-tuned model weights and can deploy anywhere without restrictions.

How do I switch between standard and reasoning modes?

Cogito models are hybrid—they seamlessly support both modes. For reasoning mode, either set enable_thinking=True in the tokenizer or add "Enable deep thinking subroutine." to your system prompt and prefill with <think>\n.

What makes the 671B MoE model special?

Our 671B MoE is among the strongest open models globally, matching DeepSeek v3 performance while using 60% shorter reasoning chains. It's our proof of concept for superintelligence through iterative self-improvement rather than pure scale.

Do you support tool calling and function calling?

Yes, all Cogito models support tool calling (single, parallel, multiple) in both standard and reasoning modes. They're compatible with standard function calling protocols.

How do I migrate from OpenAI to Cogito?

Migration is seamless with Cogito's OpenAI-compatible API on Together AI. Simply change the base URL and model name in your existing code. Same API format, better reasoning, and transparent costs with no surprises on your bill.

What are the current pricing rates for Cogito models?

Cogito 70B: $0.88 per million tokens • Cogito 109B MoE: $0.18 input / $0.59 output per million tokens • Cogito 405B: $3.50 per million tokens • Cogito 671B MoE: $1.25 per million tokens. All models offer significant cost savings compared to closed models while delivering superior reasoning efficiency.

Experience Self-Improving Intelligence

Try Cogito models now and see the difference that true iterative self-improvement makes.

Try now

Subscribe to newsletter

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.