Inference
EN
Sarvam

Sarvam 105B

Sarvam 105B is a reasoning model from Sarvam. 15 benchmarks count toward its index, in 6 categories.

  • Open weights
  • Reasoning
  • Input: text
  • Released 2026-03
RankedEnough results for an index.
IndexAll categories, out of 100.36.7±5.7Median 51.0
CoverageIndex weight that has results.85%
SpeedOutput tokens per second.191/sMedian 67/s
First tokenSeconds to the first output token.12.1 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.—Median $0.60
Output / 1MUS dollars per 1M output tokens.—Median $2.55
ContextMaximum tokens in one request.128KMedian 262K
EloLMArena rating and rank.—Median 1422

Run locally

106.03B parameters
16-bit
223.7 GB
8-bit
119.3 GB
4-bit
68.5 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy fits in 128 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.2 counted36.4186 / 218
CodingCode writing and repair.1 counted34.4—
ReasoningLogic problems and puzzles.2 counted22.1222 / 227
MultimodalTasks with images and text.None counted——
KnowledgeFacts and expert knowledge.4 counted42.4181 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.2 counted27.658 / 63
MathMath problems.4 counted46.697 / 168

Results

15 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
MATH-500 Problem SetMath98.6%2 / 27
Massive Multitask Language UnderstandingKnowledge90.6%4 / 34
American Invitational Mathematics Examination 2025Math88.3%21 / 53
Massive Multitask Language Understanding ProfessionalKnowledge81.7%35 / 75
Instruction-Following EvalInstruction84.8%41 / 70
Artificial Analysis GPQA DiamondKnowledge73.8%123 / 199
Artificial Analysis Humanity's Last ExamKnowledge11.0%141 / 222
τ²-Bench Tool-Agent-User EvaluationAgentic46.8%99 / 149
Critical Physics TasksReasoning0.0%150 / 214
GPQA Diamond · BenchLMReasoning78.7%65 / 84
Harvard-MIT Mathematics Tournament February 2025Math85.8%12 / 15
Harvard-MIT Mathematics Tournament November 2025Math85.8%10 / 12
Artificial Analysis IFBenchInstruction34.4%124 / 146
BrowseCompAgentic49.5%58 / 63
Software Engineering Benchmark VerifiedCoding45.0%105 / 114
Artificial Analysis Long Context ReasoningReasoning0.0%204 / 217

Sources

BenchLMCC BY-NC 4.0

History

5 releases
Index37.1 → 36.7
23 Oct 202612 Nov 2026

Questions

How good is Sarvam 105B?
Its index is 36.7 out of 100. That is place 230 of 283 ranked models.
What is Sarvam 105B best at?
Math: 46.6 out of 100, place 97 of 168.
How fast is Sarvam 105B?
191 output tokens per second, 12.1 s to the first token.
What is the context window of Sarvam 105B?
128K tokens.
Is Sarvam 105B open source?
Yes. Its weights are published.
When was Sarvam 105B released?
March 2026.

Badge

Sarvam 105B on Inference360: place 230 of 283, index 36.7
HTML
<a href="https://inference360.ai/models/sarvam-105b"><img src="https://inference360.ai/badges/sarvam-105b.svg" alt="Sarvam 105B on Inference360: place 230 of 283, index 36.7" height="20"></a>
Markdown
[![Sarvam 105B on Inference360: place 230 of 283, index 36.7](https://inference360.ai/badges/sarvam-105b.svg)](https://inference360.ai/models/sarvam-105b)

Compare with

Ling 2.6 Flash37.0Nemotron 3.5 Lightning 30B A3B NVFP437.0QwQ-32B36.3Sarvam 30B30.3

More from Sarvam

All models
Sarvam 30B30.3