d
WE ARE LINGO
2nd floor, Garuda BHIVE, BMTC Complex,
Old Madiwala, Kuvempu Nagar,
Stage 2, BTM Layout, Bengaluru,
Karnataka – 560068. India.
Shakti

A Breakthrough in On-Device AI

  • Shakti is a family of compact language models designed by SandLogic, ranging from 100 million to 8 billion parameters.
  • These models are optimized for efficient Natural Language Processing (NLP) applications in both edge and cloud environments.
  • Shakti models are built for real-time applications, supporting industries such as finance, healthcare, legal, retail, and e-commerce with use cases ranging from chatbots and customer service automation to complex business analytics.
  • By integrating advanced techniques such as Block Sparse Attention, Variable Grouped Query Attention (VGQA), and Rotary Positional Embeddings (RoPE),
  • Shakti effectively balances performance and efficiency for various enterprise use cases.
Hugging Face

SHAKTI MODEL - SERIES

Medical devices

01

Smart Wearables

02

Consumer Electronics

03

Switches & Routers

04

Drone with Edge AI

05

Transformer Architectures, Small Language Models, and On-Device AI

The Transformer architecture revolutionized Natural Language Processing (NLP) by leveraging the self-attention mechanism, which allowed for parallel computation and greater scalability. This innovation paved the way for Large Language Models (LLMs) to achieve state-of-the-art performance across tasks such as text generation, translation, and question answering.

  • Evolution of Transformer Architectures: Illustrate efforts to optimize Transformer architectures for real-time applications on devices with limited memory and processing power.
  • Small Language Models (SLMs): Model pruning removes parameters that minimally impact performance, while quantization reduces the precision of weights and activations, thereby lowering memory usage and computation time.
  • Advances in On-Device AI: Shakti’s architecture integrates key techniques from LLaMA and Mistral, while introducing innovations such as Variable Grouped Query Attention (VGQA) and SwiGLU activations, allowing it to deliver real-time performance on edge devices without extensive hardware.

Discover the groundbreaking innovations behind our patented technologies.

Know More

Architecture of Shakti-LLM

Shakti-LLM is optimized for resource-constrained environments like smartphones, wearables, and IoT systems. With 2.5 billion parameters and a context length of 4096 tokens, it focuses on high-performance NLP for real-time applications.

Key innovations include:

  • Variable Grouped Query Attention (VGQA): Optimizes memory usage while maintaining performance parity with conventional attention mechanisms, ideal for resource-constrained environments.
  • SwiGLU activations: Ensures stable training and inference in resource-constrained environments.
  • Rotary Positional Embeddings (RoPE): Efficiently processes long sequences for improved context understanding.
  • Block Sparse Attention: Enables efficient processing of large-scale data.
  • HALUMON Framework: Proprietary hallucination detection and mitigation system.

Shakti is designed for scalability, efficiency, and adaptability, making it suitable for industries like healthcare, finance, and customer service. It supports sliding window attention and Key-Value Caching for efficient long-sequence processing in edge computing environments.

Model Configurations For Various On-device and On-prem Apps

Shakti-100M: Edge Computing Powerhouse

Ideal for:

  • Real-time customer interaction systems
  • IoT device integration
  • Mobile application enhancement
  • Edge analytics implementation

Benchmarking performance:

  • 61.97% accuracy on PiQA
  • 51.34% on Hellaswag
  • 74.59% on reasoning tasks

Shakti-250M: Domain Intelligence Enabler

Hugging Face

Specialized Training:

  • Medical datasets including MIMIC-III and PubMed
  • Financial filings and market data
  • Legal precedent databases

Perfect for:

  • Healthcare data analysis and patient care support
  • Financial document processing and risk assessment
  • Legal document analysis and compliance monitoring
  • Automated customer service systems

Benchmarking performance:

  • 41.25% on MedQA (outperforming 1B+ models)
  • 52.97% on WinoGrande
  • 60.3% on domain-specific tasks

Shakti-500M: Enterprise Workflow Optimizer

Hugging Face

Technical Features:

  • Advanced attention mechanisms for complex queries
  • Multilingual support for 50+ languages
  • Real-time document processing capabilities

Suited for:

  • Multi-language business operations
  • Complex document processing workflows
  • Enterprise knowledge management
  • Advanced customer support systems

Performance Metrics:

  • 45.53% on Hellaswag
  • 39.80% on OpenBookQA
  • 71.40% on domain-specific tasks

Shakti-1B: Multimodal Intelligence Hub (Coming Soon...)

Data Processing Capabilities:

  • Integrated text and image understanding
  • Chart and graph analysis
  • Complex document structure comprehension

Specialized Features:

  • Medical imaging analysis
  • Financial chart interpretation
  • Technical document parsing

Enterprise Use Cases:

  • Radiology report automation
  • Financial statement analysis
  • Technical documentation processing
  • Multimedia content management

Shakti-2.5B: Enterprise Scale Solution

Hugging Face

Designed for:

  • Large-scale data analysis systems
  • Complex decision support platforms
  • Advanced multilingual applications
  • Real-time enterprise analytics
  • Enterprise knowledge management
  • Automated reporting systems

Advanced Capabilities:

  • 86.2% accuracy on PiQA
  • 76.7% on WinoGrande
  • 79.2% on SocialQA

Technical Innovations:

  • Sliding window inference for real-time processing
  • Enhanced context understanding up to 32K tokens (ROPE)
  • Advanced multilingual processing

Shakti-5B: Analytics Powerhouse (Coming Soon...)

Core Strengths:

  • Block Sparse Attention for efficient
  • large-scale processing
  • Advanced market prediction capabilities
  • Complex pattern recognition

Specialized for:

  • Market analysis and prediction systems
  • Supply chain optimization platforms
  • Advanced customer analytics
  • Complex pattern recognition applications
  • Financial market modeling

Shakti-8B: Advanced Research Platform (Coming Soon...)

Technical Specifications:

  • Extended context processing
  • Advanced multimodal integration
  • Complex reasoning capabilities

Engineered for:

  • Scientific research and analysis
  • Complex data modeling systems
  • Advanced hypothesis testing
  • Large-scale knowledge synthesis

Key Advantages

  • Flexible Deployment Across Cloud and Edge Environments
  • Advanced Quantization Support for Optimal Resource Usage
  • Comprehensive Responsible AI Framework
  • Extensive Domain Adaptation Capabilities
  • Robust Security and Monitoring Features
  • Multilingual and Multimodal Excellence
  • Scalability for Enterprises of All Sizes
  • Energy-Efficient and Eco-Friendly
  • Future-Proofed with Continuous Updates

SHAKTI MODEL - SERIES

Fairness, Accountability, and Transparency

1. Data Filtering: Remove sensitive information; add synthetic data to ensure fairness. 2. Compliance: Adhere to ethical guidelines, especially in sensitive sectors. 3. Audits: Ongoing audits and assessments for bias mitigation.

Data Preparation & Quality Control

1. Filtering & Augmentation: Remove sensitive data; fill gaps with synthetic data. 2. Enterprise Data Mapping: Ensure Shakti outputs align with specific business needs. 3. Multi-Domain Data: Incorporate diverse data types for well-rounded understanding.

LLMops & Lifecycle Management

1. Human-in-the-Loop: Continuous quality assurance and adjustments. 2. Continuous Monitoring: Retrain and update model using new data. 3. Version Control: Ensure traceable, consistent deployment. 4. Scalable Pipelines: Automate data, training, and deployment processes.

Training Methodologies

1. Supervised Fine-Tuning (SFT) Improve context-based responses in industries like healthcare and finance. 2. RLHF: Enhance user alignment and helpfulness in Shakti-500м. 3. DPO: Efficient user-aligned output tuning in Shakti-100M & Shakti- 250M. 4. In-House Training Framework: Optimize and scale model training for enterprise use.

Evaluation & Quality Metrics

1. DeepEval: Open-source tool to measure faithfulness and perplexity. 2. Common Metrics: BLEU, ROUGE, FI, Precision, Recall, BERT, MMLU. 3. Ground Truth Analysis: Verify outputs against reference data. 4.LLMHarness (Open Source) Evaluate reasoning and factual accuracy. 5. HaluMon: Detect and minimize hallucinations

Guardrails & Safety Mechanisms

1. HaluMon: Monitor hallucinations for factual reliability. 2. Prompt Tuning & Query Expansion: Maintain ethical, accurate responses. 3. Prompt Testing & Monitoring: Assess prompt changes for quality assurance.

Responsible Deployment & Ethical Use

1.In-House & Open-Source Tools: Monitor outputs and ensure rellability. 2. HaluMon & LLM Harness: Maintain ethical outputs with focus on accuracy. 3. Compliance Standards: Responsible deployment, especially in sensitive sectors.

Shakti Benchmarks

Shakti-100M: Edge Computing Powerhouse

Category Benchmark Shakti-100M Boomer-634M [1] SmolLM-135M [2] SmolLM-360M [3] AMD-Llama-135m [4]
Popular Aggregate Benchmark
MMLU (5-shot)
25.56
25.23
30.2
34.4
23.02
BigBenchHard(0- Shot)
30.12
21.11
23
24.4
18.71
IFEval
24.3
22.22
15.9
19.8
22
Language Understanding
Hellaswag (5-shot)
51.34
34.08
41.2
51.8
30.48
Anli (7-shot)
21.34
27.5
–
–
30.73
Reasoning
Piqa (5-shot)
69.2
62.57
68.4
71.6
64.20
OpenbookQA (10- shot)
37.9
35.76
34
37.2
30.73
Truthfulqa (MC2) (10-shot)
29.2
27.57
–
–
22.56
WinoGrande (5-shot)
61.3
51.07
51.3
52.8
50.12
ARC Challenge
45.8
62.57
42.4
50.1
43.64
SQuAD
31.5
57.5
–
–
25
MedQA
28.3
14
11.02
12.36
15.57
GPQA
14.9
12.1
9.89
11
12.4
Bool Q
29.4
22.9
17.3
21.3
23.54
SocialQA
23.4
14.5
16.9
19
19.1
CommonsenseQA
35.8
29
32.7
35.3
22.56
Factual Knowledge
Trivia QA (5-shot)
15.3
2.73
4.3
9.1
7.54
Math
GSM8K (8 shot)
9.2
0.91
1
1.6
–
MATH (O shot)
13.9
23.38
14
19
20.64
Coding and Programming
Humaneval
7.8
–
–
–
5.1

Shakti-250M: Domain Intelligence Enabler

Category Benchmark Shakti-250M Boomer-1B [1] Boomer-634M [2] Qwen2.5-0.5B [3] SmolLM-360M [4] Llama 3.2 1B [5]
Popular Aggregate Benchmark
MMLU (5-shot)
28.98
25.92
25.23
47.5
34.4
32.2
BigBenchHard(0- Shot)
13.75
28.65
21.11
20.3
24.4
30.93
IFEval
12.83
23.81
22.22
27.9
19.8
59.5
Language Understanding
Hellaswag (5-shot)
29.96
31.66
34.08
52.1
51.8
41.2
Anli (7-shot)
33.40
32.57
27.5
26.85
–
22.56
Reasoning
Piqa (5-shot)
63.22
60.78
62.57
72.50
71.6
80.64
OpenbookQA (10- shot)
16.60
22.56
35.75
30.73
37.2
37
Truthfulqa (MC2) (10-shot)
20.69
25.69
27.57
40.2
–
30.7
WinoGrande (5-shot)
52.97
45.79
51.07
56.3
52.8
60
ARC Challenge
41.20
22.35
62.57
35.6
50.1
32.8
SQuAD
23.25
67
57.5
52.94
–
49.2
Factual Knowledge
Trivia QA (5-shot)
1.68
25.25
2.73
12.5
9.1
25.69
Math
GSM8K (8 shot)
2.35
1.5
0.91
41.6
–
44.4
MATH (O shot)
21.71
–
23.38
19.5
–
–

Shakti-500M: Enterprise Workflow Optimizer

Category Benchmark Shakti-500M Boomer-1B [1] Boomer-634M [2] Qwen2.5-0.5B [3] OLMo 1B [4] Llama 3.2 1B [5]
Popular Aggregate Benchmark
MMLU (5-shot)
26.90
25.92
25.23
47.5
–
32.2
BigBenchHard(0- Shot)
29.1
28.65
21.11
20.3
–
30.93
IFEval
26.62
23.81
22.22
27.9
–
59.5
Language Understanding
Hellaswag (5-shot)
45.53
31.66
34.08
52.1
62.5
41.2
Anli (7-shot)
36.70
32.57
27.5
26.85
–
22.54
Reasoning
Piqa (5-shot)
74.59
60.78
62.57
72.50
73.7
80.64
Med MCQA (2-shot)
21.61
17.56
37.50
42.5
–
37.57
OpenbookQA (10-shot)
39.80
22.56
35.76
30.73
46.4
37
WinoGrande (5-shot)
59.67
45.79
51.07
56.3
58.9
60
SQuAD
71.40
67
57.5
52.94
–
49.2
Factual Knowledge
Trivia QA (5-shot)
31.11
25.25
27.6
12.5
–
25.69
Math
GSM8K (8 shot)
20.92
1.5
0.91
41.6
–
44.4
MATH (O shot)
26.97
–
23.38
19.5
–
–

Shakti-2.5B: Enterprise Scale Solution

Category Benchmark Sandlogic SLM PHi-3-Mini-4k [1] PHi-3.5-Mini [2] Sarvam-V1 [4] Llama 3 8B [3]
Popular Aggregate Benchmark
MMLU (5-shot)
71.7
68.8
69
36.2
66.5
BigBenchHard(0- Shot)
58.2
71.7
69
31.2
51.5
IFEval
75.4
59.2
–
–
77.4
Language Understanding
Hellaswag (5-shot)
52.4
76.7
69.4
48
71.1
Anli (7-shot)
55.2
52.8
49.2
32
57.3
Reasoning
Piqa (5-shot)
86.2
84.2
81
78
75.7
OpenbookQA (10- shot)
36.8
83.2
79.2
31.4
82.6
Truthfulqa (MC2) (10-shot)
68.4
65
64
48
63.1
WinoGrande (5-shot)
76.7
70.8
68.5
67
65
ARC Challenge
69.3
78.6
–
–
87.4
SQuAD
63.5
–
–
–
67.7
MedQA
60.3
53.8
50.6
38.1
60.5
GPQA
31.2
32.8
–
–
31.9
Bool Q
61.10
77.6
78
59.7
80.9
SocialQA
79.2
76.6
74.7
62
73.9
Factual Knowledge
Trivia QA (5-shot)
58.2
64
61.5
48.6
67.7
Math
GSM8K (8 shot)
72.2
77.7
–
–
86.2
MATH (O shot)
46.7
48
–
–
44.2
Coding and Programming
Humaneval
35.8
59.1
68.2
37.1
60.4