Sarvam 105B
Sarvam's flagship reasoning model, trained from scratch in India under the IndiaAI mission. Mixture-of-Experts Transformer with Multi-head Latent Attention and 128 experts; launched 18 Feb 2026 at the India AI Impact Summit and open-sourced on 6 Mar 2026 (Apache-2.0). Powers the Indus assistant.
Measured results · Artificial Analysis
Independent evaluations of Sarvam 105B (High configuration). Values are reported as published; nothing here is estimated or derived. Snapshot 2026-11-15.
- Intelligence Index
- —
- Cost to run the Index
- —
- Time per task
- —
- Omniscience index
- —
- Reasoning model
- Yes
- Cached-input price
- $0.026 / 1M
Reasoning & knowledge
- CritPt0.0%
- GPQA Diamond73.8%
- Humanity's Last Exam11.0%
- IFBench34.4%
Coding & terminal agents
Agentic work
- τ²-Bench Telecom46.8%
Long context & documents
Sourced Benchmark Evaluations (7)
Verified performance across authoritative benchmarks with provenance tracking.| Benchmark | Status | Score | Trust | Date | Source Type | Provenance |
|---|---|---|---|---|---|---|
| GPQA | Nearing Saturation | 73.8% | 78 | 2026-11-15 | independent | Source ↗ |
| Humanity's Last Exam | Active | 11% | 93 | 2026-11-15 | independent | Source ↗ |
| IFBench | Active | 34.4% | 93 | 2026-11-15 | independent | Source ↗ |
| τ²-Bench Telecom | Active | 46.8% | 93 | 2026-11-15 | independent | Source ↗ |
| Terminal-Bench Hard | Active | 1.5% | 93 | 2026-11-15 | independent | Source ↗ |
| AA-LCR (Long Context Reasoning) | Active | 0% | 93 | 2026-11-15 | independent | Source ↗ |
| CritPt | Active | 0% | 93 | 2026-11-15 | independent | Source ↗ |
# 1. Pull and execute model locally with Ollama
ollama run sarvam-105b
# 2. Or invoke via local OpenAI-compatible endpoint
curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "sarvam-105b",
"messages": [{"role": "user", "content": "Analyze reasoning chains on AIME 2026"}]
}'Evaluation & Sourcing Notes
Scores listed for Sarvam 105B represent verified evaluations extracted from official research papers, independent evaluation suites (HELM, LMSYS, OpenCompass, LiveBench), and verified audit reports.
All benchmarks marked as saturated or deprecated reflect historical performance where the benchmark no longer provides active discriminative power.