Model
Model explorer

Sarvam-105B

OPEN
Sarvam AI · Sarvam family · released Feb 18, 2026

India's first sovereign 100B+ open model, trained from scratch under the IndiaAI Mission; Apache-2.0 MoE with Multi-head Latent Attention and 128K context via YaRN, covering 22 Indian languages plus English.

ReasoningCodingVisionFunction callingTool useAgentic
1676.2
Elo · rank #90
Parameters
105B
Active params
10.3B (MoE)
Context
128K tokens
Architecture
105B MoE transformer (128 experts, ~10.3B active/token), Multi-head Latent Attention
License
Apache 2.0
Languages
23+
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
AIMEMath88.3%#51
best: GPT-5.2 · 100.0%
Arena-Hard v2Human preference71.0%#5
best: Qwen3-Max · 86.1%
BrowseCompAgents49.5%#31
best: Kimi K3 · 91.2%
GPQA DiamondReasoning78.7%#87
best: GPT-5.6 · 94.6%
Humanity's Last ExamReasoning10.1%#89
best: Claude Sonnet 5 · 57.4%
IFEvalReasoning84.8%#65
best: Gemma 4 26B A4B · 98.5%
LiveCodeBenchCoding71.7%#65
best: DeepSeek-V4-Pro (Think Max) · 93.5%
MATH-500Math98.6%#3
best: GPT-5 · 99.4%
MMLU-ProKnowledge81.7%#49
best: Claude Fable 5 · 91.5%
MMLUKnowledge90.6%#7
best: OpenAI o3 · 92.9%
SWE-bench VerifiedCoding45.0%#97
best: Claude Fable 5 · 95.0%
Run it locally
VRAM @ Q4
65 GB
VRAM @ FP16
—
Fits on (Q4)
M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
Too large to run on a single RTX 4090 even at Q4 (~64GB needed); no consumer single-GPU llama.cpp benchmark found.
Quantizations
GGUF Q4 · FP8
Sarvam family
Elo progression across releases
Sarvam-1 (2B)Oct 2024129.0Sarvam-105BFeb 20261676.2
Resources
HFsarvamai/sarvam-105bDOCSsarvam.aiAPIAvailable via Sarvam AI API
API price weights · each benchmark row carries its own source badge (see methodology)