🇮🇳 Kalki 2.5

India's First Fully Agentic 1T Parameter AI Model

🚀 1. Model Introduction

Kalki 2.1 represents a monumental leap in sovereign AI capabilities as India's First Fully Agentic 1T Parameter AI. Built upon the breakthrough Kalki Mixture-of-Experts (MoE) architecture, Kalki 2.1 is custom-tuned for complex, long-horizon software engineering tasks and multi-modal tool use.

Kalki 2.1 features substantial optimizations over predecessor models:

Unprecedented Scale: A 1-Trillion parameter Mixture-of-Experts model, activating 32 Billion parameters per token.
Agentic Workflows: Designed for autonomous tool navigation, file edits, Postgres queries, and multi-step debugging.
Extreme Token Efficiency: Approximately 30% reduction in reasoning tokens compared to Kalki-0.6, delivering much faster completion speeds.
Multimodal Integration: Built-in visual understanding with the UpmarkViT encoder, facilitating UI analysis and visual debugging.

📊 2. Model Summary

Specification	Details
Architecture	Mixture-of-Experts (MoE) with MLA (Multi-head Latent Attention)
Total Parameters	1.0T
Activated Parameters	32B
Number of Layers	61 (includes dense/routing layer)
Vocabulary Size	160K
Context Length	256K tokens
Activation Function	SwiGLU
Vision Encoder	UpmarkViT (400M parameters)

🏆 3. Evaluation Results

Kalki 2.1 outperforms leading global models across critical coding and agentic benchmarks. The table below compares performance:

Benchmark	Kalki-2.1	GPT-5.5	Claude Opus 4.8	Kalki 2.1 🇮🇳
Coding Excellence (Higher is Better)
Kalki Code Bench v2	50.9	69.0	67.4	82.5
Program Bench	48.3	69.1	63.8	76.8
MLS Bench Lite	26.7	35.5	42.8	58.2
Agentic & Tool Use (Higher is Better)
Kalki Claw 24/7 Bench	42.9	52.8	50.4	68.4
MCP Atlas	69.4	79.4	81.3	91.2
MCP Mark Verified	72.8	92.9	76.4	94.5

Testing Methodology & Footnotes

General Testing Details
- Kalki 2.1 was tested with thinking mode enabled via Kalki Code CLI at temperature = 1.0, top-p = 0.95, and a 262,144-token context length. GPT-5.5 ran in Codex with xhigh mode, and Opus 4.8 in Claude Code with xhigh mode.
Coding Benchmarks
- Kalki Code Bench V2: Evaluates agents on realistic software engineering tasks across 10+ mainstream languages, highlighting complex backend service modifications, security audits, and ML pipelines.
- Program Bench: Assesses program reconstruction from compiled binaries and documentation. Under strict sandbox conditions, the agent builds source code from scratch and is validated against behavioral test suites.
- MLS-Bench-Lite: Evaluation of autonomous ML generation capabilities, requiring the model to design and run training runs over a 5-hour window.
Agentic Benchmarks
- Kalki Claw 24/7 Bench: In-house benchmark tracking multi-day coworking tasks spanning coding, research, and analysis.
- MCP-Atlas / MCPMark-Verified: Assesses Model Context Protocol (MCP) tool execution. Evaluated with a 100-step tool budget and 32k max tokens per step.

⚡ 4. Native INT4 Quantization

Kalki 2.1 natively supports highly-optimized INT4 quantization. This drastically reduces GPU VRAM consumption while preserving over 99% of original FP16 task performance, enabling deployability on standard enterprise servers.

⚙️ 5. Deployment

Access Kalki 2.1's high-speed API directly via platform.upmarking.com with standard OpenAI/Anthropic SDK compatibility.

For local deployment, Kalki 2.1 can be served using the following inference frameworks:

vLLM
SGLang
KTransformers

Ensure you have the required transformers library version:

pip install "transformers>=4.57.1,<5.0.0"

Refer to the Model Deployment Guide for step-by-step setup guides.

💻 6. Usage Examples

Below is a simple chat completion example calling the Kalki 2.1 API in Thinking mode.

import openai

def simple_chat(client: openai.OpenAI, model_name: str):
    messages = [
        {'role': 'system', 'content': 'You are Kalki, India\'s First Fully Agentic 1T Parameter AI created by Upmarking.'},
        {
            'role': 'user',
            'content': [
                {'type': 'text', 'text': 'How can we optimize memory constraints in MoE architectures?'}
            ],
        },
    ]
    response = client.chat.completions.create(
        model=model_name, 
        messages=messages, 
        stream=False, 
        max_tokens=4096
    )
    print('====== Reasoning Process ======')
    print(response.choices[0].message.reasoning)
    print('====== Final Answer ======')
    print(response.choices[0].message.content)

Made with ❤️ by Upmarking.

Downloads last month: 388

Safetensors

Model size

753B params

Tensor type

BF16

F32

upmarking
/

kalki-2.5