Cogito v2.1 671B

Advanced hybrid reasoning model with self-improving capabilities

About model

Cogito v2.1 671B is Deep Cogito's flagship open-source hybrid reasoning model, built on Iterated Distillation and Amplification (IDA) that learns to think better through self-improvement. Outperforming all US open models and rivaling Claude 4 Opus and O3, Cogito v2.1 achieves frontier-level performance while using 60% shorter reasoning chains than competitors — delivering breakthrough efficiency with 4,894 average tokens per response (lowest among frontier models) at just $1.25 per million tokens.

AIME 2025 (Competition Math)

89.47%

Elite mathematical reasoning outperforming models 10x larger

More Efficient Reasoning

60%

Shorter chains than DeepSeek R1 with equal accuracy

Avg Tokens per Response

4,894

Lowest among all frontier models for massive cost savings

Model key capabilities

Hybrid Reasoning Modes: Seamlessly switch between fast standard responses and deep step-by-step reasoning
Self-Improving Intelligence: IDA methodology distills reasoning discoveries back into parameters, compounding over time
State-of-the-Art Benchmarks: 98.57% MATH-500, 77.72% GPQA Diamond, 84.69% MMLU Pro
Production-Ready Efficiency: 128K context window, OpenAI-compatible API, native tool calling support

Quickstart guides

Building a RAG Workflow

Performance benchmarks

Model	GPQA Diamond	HLE	LiveCodeBench	MATH500	SWE-bench verified
Cogito v2.1 671B	76.0%
Related open-source models
Competitor closed-source models
Claude Opus 4.6	90.5%	34.2%			78.7%
OpenAI o3	83.3%	24.9%		99.2%	62.3%
OpenAI o1	76.8%			96.4%	48.9%
GPT-4o	49.2%	2.7%	32.3%	89.3%	31.0%

API usage

cURL
Python
Typescript

Endpoint:

deepcogito/cogito-v2-1-671b

curl -X POST "https://api.together.xyz/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepcogito/cogito-v2-1-671b",
    "messages": [
      {
        "role": "user",
        "content": "What are some fun things to do in New York?"
      }
    ]
}'

from together import Together

client = Together()

response = client.chat.completions.create(
  model="deepcogito/cogito-v2-1-671b",
  messages=[
    {
      "role": "user",
      "content": "What are some fun things to do in New York?"
    }
  ]
)
print(response.choices[0].message.content)

import Together from 'together-ai';
const together = new Together();

const completion = await together.chat.completions.create({
  model: 'deepcogito/cogito-v2-1-671b',
  messages: [
    {
      role: 'user',
      content: 'What are some fun things to do in New York?'
     }
  ],
});

console.log(completion.choices[0].message.content);

Model card
Architecture Overview:
• Cogito v2.1 671B employs a sophisticated Mixture-of-Experts (MoE) architecture with 671 billion total parameters, utilizing sparse routing mechanisms to activate only specialized expert subnetworks per token, enabling massive scale without proportional compute costs
• Features a 128K token context window optimized for long-form reasoning, technical documentation, and multi-turn conversations
• Implements a hybrid inference system supporting both standard mode (direct answers using internalized "intuition") and reasoning mode (step-by-step self-reflection with visible thought chains)
• Optimized for efficient serverless deployment on Together AI's infrastructure

Training Methodology - Iterated Distillation & Amplification (IDA):
• Revolutionary self-improvement approach where the model runs reasoning chains during training, then is trained on its own intermediate thoughts to develop stronger "machine intuition"
• Unlike traditional models that rely on extended inference-time reasoning, Cogito distills successful reasoning patterns directly into model parameters
• Training process explicitly rewards shorter, more efficient reasoning paths while discouraging unnecessary computational detours
• Trained on multilingual datasets spanning 30+ languages with emphasis on coding, STEM, instruction following, and general helpfulness
• Total training cost remarkably achieved at under $3.5 million for the entire Cogito family (3B to 671B), demonstrating unprecedented cost efficiency

Performance Characteristics:
• AIME 2025 (Competition Mathematics): 89.47% - outperforming models 10x larger
• MATH-500 benchmark: 98.57% accuracy
• GPQA Diamond (Scientific Reasoning): 77.72%
• SWE-Bench Verified (Coding): 42.00% solve rate
• MMLU Pro (Reasoning & Knowledge): 84.69%
• Multilingual MMLU: 86.24% across 30+ languages
• Average token efficiency: 4,894 tokens per response (lowest among frontier models)
• Performance competitive with DeepSeek v3, matching or exceeding latest 0528 model while using 60% shorter reasoning chains
• Approaches capabilities of closed models like Claude 4 Opus, O3, and GPT-5 across diverse benchmarks
• Demonstrates emergent multimodal reasoning capabilities, able to reason about images despite not being explicitly trained for visual tasks
‍
Applications & use cases
High-Performance Use Cases:
• Advanced Mathematical Problem Solving: Superior performance on competition mathematics (AIME 2025: 89.47%), calculus, optimization problems, and quantitative analysis
• Software Engineering & Code Generation: 42% solve rate on SWE-Bench demonstrates strong debugging, code review, and system design capabilities
• Scientific Research & STEM: 77.72% on GPQA Diamond showcases expertise in physics, chemistry, biology, and interdisciplinary scientific reasoning
• Multilingual Applications: 86.24% on Multilingual MMLU enables global deployment across 30+ languages with native-level comprehension
• Legal & Policy Analysis: Reasoning mode excels at applying precedents, analyzing case law, and providing nuanced legal interpretations

Enterprise Applications:
• Intelligent Document Processing: 128K context window handles entire technical documents, contracts, research papers in single context
• Customer Support Automation: Hybrid mode allows fast responses for simple queries, deep reasoning for complex troubleshooting
• Financial Analysis & Risk Assessment: Strong quantitative reasoning combined with efficient token usage for cost-effective at-scale deployment
• Educational Technology: Step-by-step reasoning mode ideal for tutoring, homework help, and adaptive learning systems
• Research Assistance: Frontier performance at $1.25/1M tokens makes large-scale research analysis economically viable

Developer & Research Applications:
• Rapid Prototyping: Together AI's serverless platform enables instant deployment without infrastructure setup
• Model Experimentation: Compare standard vs reasoning modes in real-time via playground interface
• Benchmark Development: Performance approaching closed frontier models enables reproducible research
• Scalable Research: Serverless infrastructure scales automatically for large-scale experiments

Cost-Sensitive Deployments:
• High-Volume Production: Lowest token usage (4,894 avg) among frontier models translates to 20-40% cost savings vs alternatives
• Serverless Efficiency: Pay-per-use pricing on Together AI eliminates infrastructure costs and management overhead
• Startup & SMB Applications: Frontier capabilities at accessible pricing ($1.25/1M tokens) democratizes advanced AI
• Auto-scaling: Together AI's serverless infrastructure automatically handles traffic spikes without manual intervention

Unique Capabilities:
• Emergent Image Reasoning: Despite no explicit visual training, demonstrates ability to reason about images when presented in context
• Efficiency-First Design: 60% shorter reasoning chains mean faster responses and lower costs without sacrificing accuracy
• Hybrid Intelligence: Seamlessly switch between fast intuition and deep deliberation based on query complexity
‍

Related models

Model specifications

Model data

Model provider
Deep Cogito
Type
LLM
Chat
Main use cases
Chat
Small & Fast
Deployment
Serverless
Endpoint
deepcogito/cogito-v2-1-671b
Parameters
671B (MoE)
Context length
32K
Input price
$1.25 / 1M tokens
Output price
$1.25 / 1M tokens
Input modalities
Text
Output modalities
Text

Quantization level
FP8
Category
Chat

Run in Playground

Quickstart docs

Deploy model

Cogito v2.1 671B

About model

API usage

Model card

Applications & use cases