Sarvam-105B
Active Parameters
106B
Context Length
128K
Modality
Text
Architecture
Mixture of Experts (MoE)
License
Apache 2.0
Release Date
6 Mar 2026
Knowledge Cutoff
-
System Requirements
VRAM requirements for different quantization methods and context sizes
1,024 tokens
Consumer
11x RTX 4090
24GB VRAM
Datacenter
4x NVIDIA A100
80GB VRAM
Apple Silicon
3x Apple M3 Max
128GB VRAM
128,000 tokens
Consumer
16x RTX 4090
24GB VRAM
Datacenter
5x NVIDIA A100
80GB VRAM
Apple Silicon
3x Apple M3 Max
128GB VRAM
Architecture Diagram
Evaluation Benchmarks
No evaluation benchmarks for Sarvam-105B available.
Rankings
Overall Rank
-
Coding Rank
-
About Sarvam-105B
Sarvam-105B is an advanced Mixture-of-Experts (MoE) model with 106B total parameters and 10.3B active parameters, designed for superior performance across complex tasks. Released March 6, 2026 under Apache 2.0 license. Uses MLA-style attention stack with decoupled QK head dimensions (q_head_dim=192, v_head_dim=128), large head_dim of 576, and 128 experts with top-8 routing. Features 128K native context (extensible via YaRN scaling with factor 40), and delivers exceptional performance in agentic tasks, mathematics, and coding. Consistently matches or surpasses major closed-source models with state-of-the-art results across 22 Indian languages while maintaining competitive global benchmark performance.
Technical Specifications
Attention
Attention Structure
Multi-Layer Attention
Attention Heads
64
Key-Value Heads
-
Attention Head Dimension
576
Position Embedding
ROPE
RoPE Theta
10,000
Sliding Window Attention
No
Sliding Window Size
-
Sliding Window Ratio
-
Linear Attention
-
Linear Attention Ratio
-
Normalization
RMS Normalization
Activation Function
SwigLU
Dimensions
Hidden Dimension Size
4,096
Number of Layers
32
FFN Intermediate Size (Dense)
2,048
Multi-Token Prediction Heads
-
Tokenizer
Vocabulary Size
262,144
Mixture of Experts
Total Expert Parameters
10.3B
Number of Experts
128
Active Experts
8
Shared Experts
1
FFN Intermediate Size (per Expert)
2,048
Dense Layers Before MoE
1
About Sarvam
Sarvam AI's sovereign foundation models built for India's languages, culture, and context. Released in March 2026, these advanced Mixture-of-Experts (MoE) models offer state-of-the-art performance across 22 Indian languages while maintaining competitive results on global benchmarks. Designed with focus on reasoning, coding, multilingual capabilities, and agentic tasks. Open-sourced under Apache 2.0 license, optimized for practical deployment from resource-constrained environments to high-performance applications.