Sarvam AILanguage model
Sarvam 105B
Sarvam 105B has no overall index value yet because there are too few independent benchmark results.
Intelligencenot ranked yet
n/aOverall indexhigher is better
SpeedRank 36 of 101
107Tokens per secondhigher is better
Priceno API price
n/aper 1M tokenslower is better
ContextRank 94 of 103
131Ktokens at once≈ 200 A4 pages
PopularityRank 526 of 785
31out of 100-1 vs. last month
based on search, usage and attention
Summary
At 107 tokens per second it writes faster than a typical model (87).
Sarvam 105B is Sarvam AI’s large open Mixture-of-Experts model for Indian languages, agentic work, and demanding reasoning. The model card lists 105 billion total parameters and 10.3 billion active parameters. MLA and YaRN support its documented 128,000-token context. Sarvam emphasizes coding, mathematics, tool use, and work across 22 Indian languages. The weights are released under Apache 2.0.
Sarvam 30B is the smaller family partner aimed more directly at practical deployment. For Sarvam 105B, the card documents local inference but still calls for a source installation or hot patch with vLLM. Sarvam’s release post also documents API access and production use, even though only 10.3 billion of its 105 billion parameters are active.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 60 published benchmark results
Page 1 of 5
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
CritPt v2: Overall (Show measurement, test conditions, and source)
- Source value
- 0
- Score
- 0
- Metric
- pass@1 accuracy
- Unit
- %
- Benchmark version
- v2
- Category
- reasoning
- Task
- overall
- Task label
- Overall
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis
- Status
- disputed
- Retrieved at
- 2026-11-02
- Methodology
- Exact AA configuration: Sarvam 105B (High) (release sarvam-105b, row sarvam-105b, UUID 45790612-02e3-4c42-b5bd-cd7ed2ea1f2f, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 2000. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. The full CritPt v2 run covers 70 questions with five repeats. The official grading server reports pass@1.. Source model ID: artificial-analysis/45790612-02e3-4c42-b5bd-cd7ed2ea1f2f/sarvam-105b/sarvam-105b. Source model name: Sarvam 105B. Provider reported by source: Sarvam. Model snapshot: Sarvam 105B (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Tool policy: Python functions, symbolic expressions, and numerical answers; official grading server. Attempts: 70 questions, five repeats per question, pass@1. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 0.0%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-11-02-67d46be20a5b. Source payload hash: 67d46be20a5b412db6cfb60ad190441fcaa5513500b8fd47d033168d0e97b6b2. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-11-02. Source cohort: artificial-analysis-critpt-v2:v2:standalone:reasoning:effort=high:suffix=High
- Comparison methodology
- Exact AA configuration: Sarvam 105B (High) (release sarvam-105b, row sarvam-105b, UUID 45790612-02e3-4c42-b5bd-cd7ed2ea1f2f, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 2000. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. The full CritPt v2 run covers 70 questions with five repeats. The official grading server reports pass@1.. Tool policy: Python functions, symbolic expressions, and numerical answers; official grading server. Attempts: 70 questions, five repeats per question, pass@1. Reasoning: reasoning enabled. Reasoning effort: high
- Context
- This AA source field contains the original CritPt score. The earlier CritPt v2 attribution was incorrect; the value and former label remain as disputed historical evidence.. The source dataset marks this result as disputed. It remains visible as documented evidence, but is not reliable comparative proof.
IFBench: Overall (Show measurement, test conditions, and source)
- Source value
- 34.3537414965986
- Score
- 34.3537414965986
- Metric
- IFBench prompt-level loose accuracy (pass@1)
- Unit
- %
- Benchmark version
- AA test set: 294 questions
- Category
- general
- Task
- overall
- Task label
- Overall
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis
- Status
- superseded
- Retrieved at
- 2026-11-02
- Methodology
- Exact AA configuration: Sarvam 105B (High) (release sarvam-105b, row sarvam-105b, UUID 45790612-02e3-4c42-b5bd-cd7ed2ea1f2f, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 2000. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. Artificial Analysis runs 294 questions from the IFBench test set with five repeats per question. The official evaluator scores prompt-level accuracy in loose mode as pass@1.. Source model ID: artificial-analysis/45790612-02e3-4c42-b5bd-cd7ed2ea1f2f/sarvam-105b/sarvam-105b. Source model name: Sarvam 105B. Provider reported by source: Sarvam. Model snapshot: Sarvam 105B (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Tool policy: Single-turn answer evaluated with the official IFBench source code. Attempts: pass@1, 5 repeats per question. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 34.4%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-11-02-67d46be20a5b. Source payload hash: 67d46be20a5b412db6cfb60ad190441fcaa5513500b8fd47d033168d0e97b6b2. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-11-02. Source cohort: artificial-analysis-ifbench:AA test set: 294 questions:standalone:reasoning:effort=high:suffix=High
- Comparison methodology
- Exact AA configuration: Sarvam 105B (High) (release sarvam-105b, row sarvam-105b, UUID 45790612-02e3-4c42-b5bd-cd7ed2ea1f2f, reasoning=true, effort=high, suffix=High). AA source value for average reasoning tokens before the answer: 2000. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. Artificial Analysis runs 294 questions from the IFBench test set with five repeats per question. The official evaluator scores prompt-level accuracy in loose mode as pass@1.. Tool policy: Single-turn answer evaluated with the official IFBench source code. Attempts: pass@1, 5 repeats per question. Reasoning: reasoning enabled. Reasoning effort: high
- Context
- The often cited 58 refers to verifiable out-of-domain constraint types, not the number of questions in AA's test run.. This result is retained as historical evidence and is not interchangeable with current results.
IFBench: Overall (Show measurement, test conditions, and source)
- Source value
- 34.3537414965986
- Score
- 34.3537414965986
- Metric
- IFBench prompt-level loose accuracy (pass@1)
- Unit
- %
- Benchmark version
- AA test set: 294 questions
- Category
- general
- Task
- overall
- Task label
- Overall
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis
- Status
- active
- Retrieved at
- 2026-11-07
- Methodology
- Exact AA configuration: Sarvam 105B (High) (release sarvam-105b, row sarvam-105b, UUID 45790612-02e3-4c42-b5bd-cd7ed2ea1f2f, reasoning=true, effort=high, suffix=High). This source capture does not provide an average reasoning-token count. Artificial Analysis runs 294 questions from the IFBench test set with five repeats per question. The official evaluator scores prompt-level accuracy in loose mode as pass@1.. Source model ID: artificial-analysis/45790612-02e3-4c42-b5bd-cd7ed2ea1f2f/sarvam-105b/sarvam-105b. Source model name: Sarvam 105B. Provider reported by source: Sarvam. Model snapshot: Sarvam 105B (High). Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Tool policy: Single-turn answer evaluated with the official IFBench source code. Attempts: pass@1, 5 repeats per question. Reasoning: reasoning enabled. Reasoning effort: high. Score display: 34.4%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-11-07-b96e3ffc8c5b. Source payload hash: b96e3ffc8c5b052aa56f96d05e89d02b8cc234a75085820fca0ca2a39e498274. Parser version: artificial-analysis-rsc-models-v4. Imported at: 2026-11-07. Source cohort: artificial-analysis-ifbench:AA test set: 294 questions:standalone:reasoning:effort=high:suffix=High
- Comparison methodology
- Exact AA configuration: Sarvam 105B (High) (release sarvam-105b, row sarvam-105b, UUID 45790612-02e3-4c42-b5bd-cd7ed2ea1f2f, reasoning=true, effort=high, suffix=High). This source capture does not provide an average reasoning-token count. Artificial Analysis runs 294 questions from the IFBench test set with five repeats per question. The official evaluator scores prompt-level accuracy in loose mode as pass@1.. Tool policy: Single-turn answer evaluated with the official IFBench source code. Attempts: pass@1, 5 repeats per question. Reasoning: reasoning enabled. Reasoning effort: high
- Context
- The often cited 58 refers to verifiable out-of-domain constraint types, not the number of questions in AA's test run.
XL-SafetyBench (India-local model) (Show measurement, test conditions, and source)
- Source value
- 34.7
- Score
- 34.7
- Metric
- Attack Success Rate (ASR)
- Unit
- %
- Benchmark version
- XL-SafetyBench; paper arXiv:2605.05662v2
- Category
- safety
- Direction
- Lower is better
- Sample
- 449
- Source type
- community-reported
- Evaluator
- XL-SafetyBench, arXiv:2605.05662v2
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Table 5, India local-model result; the source describes 450 country-specific jailbreak prompts per country. Sarvam API, max_tokens 8192, temperature 0.0, up to 3 retries on empty/transient response; OpenAI Python SDK timeout 3600 s. One Sarvam-105B empty jailbreak responses after retries was marked as errors and excluded. Other Sarvam API parameters are not reported in the source; no immutable checkpoint/API model ID is given.
- Context
- The paper identifies the Sarvam model labels and the Sarvam API route, but publishes no immutable API model/checkpoint identifier or benchmark dataset version, and no confidence intervals. The multi-institution author list alone does not establish evaluator independence.
XL-SafetyBench (India-local model) (Show measurement, test conditions, and source)
- Source value
- 1.6
- Score
- 1.6
- Metric
- Neutral-Safe Rate (NSR)
- Unit
- %
- Benchmark version
- XL-SafetyBench; paper arXiv:2605.05662v2
- Category
- safety
- Direction
- Lower is better
- Sample
- 449
- Source type
- community-reported
- Evaluator
- XL-SafetyBench, arXiv:2605.05662v2
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Table 5, India local-model result; the source describes 450 country-specific jailbreak prompts per country. Sarvam API, max_tokens 8192, temperature 0.0, up to 3 retries on empty/transient response; OpenAI Python SDK timeout 3600 s. One Sarvam-105B empty jailbreak responses after retries was marked as errors and excluded. Other Sarvam API parameters are not reported in the source; no immutable checkpoint/API model ID is given.
- Context
- The paper identifies the Sarvam model labels and the Sarvam API route, but publishes no immutable API model/checkpoint identifier or benchmark dataset version, and no confidence intervals. The multi-institution author list alone does not establish evaluator independence.
XL-SafetyBench (India-local model) (Show measurement, test conditions, and source)
- Source value
- 7
- Score
- 7
- Metric
- Cultural Sensitivity Rate (CSR)
- Unit
- %
- Benchmark version
- XL-SafetyBench; paper arXiv:2605.05662v2
- Category
- safety
- Direction
- Higher is better
- Sample
- 100
- Source type
- community-reported
- Evaluator
- XL-SafetyBench, arXiv:2605.05662v2
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Table 15, India local-model result on 100 Indian cultural scenarios. Sarvam API, max_tokens 8192, temperature 0.0, up to 3 retries on empty/transient responses; OpenAI Python SDK timeout 3600 s. CSR is conditional on contextually relevant responses; the reported 100.0% engagement implies 100 relevant responses out of 100. Other Sarvam API parameters are not reported in the source; no immutable checkpoint/API model ID is given.
- Context
- The paper identifies the Sarvam model labels and the Sarvam API route, but publishes no immutable API model/checkpoint identifier or benchmark dataset version, and no confidence intervals. The multi-institution author list alone does not establish evaluator independence.
XL-SafetyBench (India-local model) (Show measurement, test conditions, and source)
- Source value
- 100
- Score
- 100
- Metric
- Engagement rate
- Unit
- %
- Benchmark version
- XL-SafetyBench; paper arXiv:2605.05662v2
- Category
- safety
- Direction
- Higher is better
- Sample
- 100
- Source type
- community-reported
- Evaluator
- XL-SafetyBench, arXiv:2605.05662v2
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Table 15, India local-model result on 100 Indian cultural scenarios. Sarvam API, max_tokens 8192, temperature 0.0, up to 3 retries on empty/transient responses; OpenAI Python SDK timeout 3600 s. Other Sarvam API parameters are not reported in the source; no immutable checkpoint/API model ID is given.
- Context
- The paper identifies the Sarvam model labels and the Sarvam API route, but publishes no immutable API model/checkpoint identifier or benchmark dataset version, and no confidence intervals. The multi-institution author list alone does not establish evaluator independence.
XL-SafetyBench (India-local model) (Show measurement, test conditions, and source)
- Source value
- 7
- Score
- 7
- Metric
- Unconditional Cultural Sensitivity Rate (uCSR)
- Unit
- %
- Benchmark version
- XL-SafetyBench; paper arXiv:2605.05662v2
- Category
- safety
- Direction
- Higher is better
- Sample
- 100
- Source type
- community-reported
- Evaluator
- XL-SafetyBench, arXiv:2605.05662v2
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Table 15, India local-model result on 100 Indian cultural scenarios. Sarvam API, max_tokens 8192, temperature 0.0, up to 3 retries on empty/transient responses; OpenAI Python SDK timeout 3600 s. Other Sarvam API parameters are not reported in the source; no immutable checkpoint/API model ID is given.
- Context
- The paper identifies the Sarvam model labels and the Sarvam API route, but publishes no immutable API model/checkpoint identifier or benchmark dataset version, and no confidence intervals. The multi-institution author list alone does not establish evaluator independence.
XL-SafetyBench (India-local model) (Show measurement, test conditions, and source)
- Source value
- 63.7
- Score
- 63.7
- Metric
- Ideal Safe Rate
- Unit
- %
- Benchmark version
- XL-SafetyBench; paper arXiv:2605.05662v2
- Category
- safety
- Direction
- Higher is better
- Sample
- 449
- Source type
- community-reported
- Evaluator
- XL-SafetyBench, arXiv:2605.05662v2
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Table 5, India local-model result; the source describes 450 country-specific jailbreak prompts per country. Sarvam API, max_tokens 8192, temperature 0.0, up to 3 retries on empty/transient response; OpenAI Python SDK timeout 3600 s. One Sarvam-105B empty jailbreak responses after retries was marked as errors and excluded. Other Sarvam API parameters are not reported in the source; no immutable checkpoint/API model ID is given.
- Context
- The paper identifies the Sarvam model labels and the Sarvam API route, but publishes no immutable API model/checkpoint identifier or benchmark dataset version, and no confidence intervals. The multi-institution author list alone does not establish evaluator independence.
Q&A over Indian-language policies (Show measurement, test conditions, and source)
- Source value
- 96
- Score
- 96
- Metric
- Correct-answer rate on an authored Indian-language company-document RAG test
- Unit
- %
- Direction
- Higher is better
- Sample
- 132
- Source type
- independent-evaluator
- Evaluator
- Sarvam Field Test
- Status
- active
- Retrieved at
- 2026-10-22
- Methodology
- Independent evaluation. Sarvam API. The result row identifies Sarvam-105B · thinking, which the methods map to medium reasoning effort. The authored Indian-language company-document RAG test contains 132 items and is reported as run on 11 October 2026. No immutable checkpoint or API revision is stated.
- Context
- The methods state that scores use 95% bootstrap intervals with 2,000 resamples. The opened report gives no numerical bounds for this row or immutable served checkpoint/API revision.
Gödel Machines refusal-asymmetry audit: overall harmful-prompt compliance (Show measurement, test conditions, and source)
- Source value
- 3.9
- Score
- 3.9
- Metric
- Harmful-prompt compliance rate
- Unit
- %
- Category
- safety
- Direction
- Lower is better
- Sample
- 1400
- Source type
- independent-evaluator
- Evaluator
- Gödel Machines, Sarvam's illusion of safety
- Status
- active
- Retrieved at
- 2026-10-23
- Methodology
- Independent evaluation. Independent evaluator's natural harmful-prompt refusal-asymmetry test. The report says 1,400 prompts per model across 14 languages, human-translated from prompts spanning ten India-specific categories, judged by GPT-5-mini. The report does not state an exact run date or immutable weight revision/API route.
- Context
- The report identifies the evaluated system only by the named Sarvam model label; it gives no immutable weight revision or evaluation route. This is a safety observation, not a general model-quality or ranking result.
Gödel Machines refusal-asymmetry audit: Privacy violation (Show measurement, test conditions, and source)
- Source value
- 6.4
- Score
- 6.4
- Metric
- Harmful-prompt compliance rate
- Unit
- %
- Category
- safety
- Direction
- Lower is better
- Source type
- independent-evaluator
- Evaluator
- Gödel Machines, Sarvam's illusion of safety
- Status
- active
- Retrieved at
- 2026-10-23
- Methodology
- Independent evaluation. Independent natural harmful-prompt safety audit, refusal-asymmetry category Privacy violation, reported for Sarvam-105B. The paper describes 100 prompts per language across 14 languages and ten categories, but does not provide the denominator for this individual category cell; sample count is omitted. Responses were judged by GPT-5-mini. No exact run date or immutable weight revision/API route is stated.
- Context
- The report identifies the evaluated system only by the named Sarvam model label; it gives no immutable weight revision or evaluation route. This is a safety observation, not a general model-quality or ranking result.
Page 1 of 5
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
The 100 most popular models of this category, 0 to 100, higher is more popular
20 of 101 models
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 7, 2026
Sarvam 105BPixtral 12B 31Agnes 3.0 Flash Preview 30Qwen 3 VL 8B Thinking 30Yi-1.5 34B 30
Method changes in the timeline, a jump at these points is no market event:
- 1November 1, 2026: Value only with search or usage data, 30-day attention
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
Similar models
More models of the same kind as Sarvam 105B, chosen by model line, developer, size, price and release date.
All specifications
Architecture
Input and output
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
Frequently asked questions about Sarvam 105B
How large is the context window of Sarvam 105B?
131,072 tokens. That is how much text Sarvam 105B reads at once, about 200 A4 pages.
Page count as a rule of thumb from Artificial Analysis
GLM-5.1200K
DeepSeek-V3 0324164K
Sarvam 105BRank 94 of 103131K
GLM-4.5131K
GPT OSS 120B131K
DeepSeek-V3.2128K
DeepSeek-R1128K
How fast is Sarvam 105B?
107 tokens per second. That is faster than a typical model at 87.
Measured by Artificial Analysis on Nov 6, 2026. AA 3.0.0, including time to first token, averaged across available workloads and five days.
GPT-6 Sol111
o3110
o4-mini110
Sarvam 105BRank 36 of 101107
Claude Opus 5.5106
Claude Sonnet 5103
Grok 4.20 Reasoning102
Can I run Sarvam 105B locally?
Yes, the weights are open. With 105B parameters it needs a lot of memory, more a server than a laptop.
Stated by Sarvam AI
- Open weights
- yes
- License
- Apache 2.0
- Parameters
- 105B
- active per token
- n/a
Sources and data date
Show all 20 sources
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
| Evaluator | Test series | Composite indices | Recorded active measurements |
|---|---|---|---|
| Artificial Analysis | 1 | 0 | 1 |
Without vendor claims and historical or disputed results; those stay labelled in the single values.
| Type | Evidence and data date |
|---|---|
| Speed | Sarvam, High effort. AA 3.0.0, including time to first token, averaged across available workloads and five days. Retrieved November 6, 2026. |
| Input | Documented modalities. Retrieved October 27, 2026. |
| Output | Documented modalities. Retrieved October 27, 2026. |
| Research status | Research date October 15, 2026. 9 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | score-units-and-evaluation-details Note The card leaves table score units, most benchmark versions, task dates, sampling and prompting unspecified. AIME's two parenthetical tool variants were excluded because their mapping is not clear from the table. |
| Not published | evaluation-methodology Note The release page omits evaluation dates, per-task denominators, score scales for several tables, and most prompt, decoding, and tool-scaffold details. The AIME table row and prose leave the first value's tool status unclear. |
| Not published | tested-weight-or-serving-revision Note The evaluator gives the public model ID and visible setup details, but no immutable tested weight revision or serving/API revision. |
| Not published | confidence-interval-95 Note No confidence interval is published for the aggregate 200-problem score. |
| Not published | underlying-run-date Note The page update timestamp is visible, but no exact date for the model's underlying evaluation run is stated. |
| Additional source | Hugging Face (retrieved October 15, 2026) · Sarvam 105B model card · Editorial description reviewed October 2, 2026 |
| Additional source | Artificial Analysis (retrieved November 2, 2026) |
| Additional source | Artificial Analysis (retrieved November 2, 2026; November 7, 2026) |
| Additional source | Sarvam: Open-Sourcing Sarvam 30B and 105B (retrieved October 17, 2026; October 23, 2026) · Sarvam 30B and 105B release post · Editorial description reviewed October 2, 2026 |
| Additional source | MathBench.ai exact-model result page (retrieved October 17, 2026; October 28, 2026) |
| Additional source | AgentAudit paper, arXiv:2609.09875 (retrieved October 17, 2026) |
| Additional source | XL-SafetyBench, arXiv:2605.05662v2 (retrieved October 18, 2026) |
| Additional source | Sarvam AI, Everything we announced at Sarvam Epoch (retrieved October 20, 2026) |
| Additional source | Sarvam Field Test (retrieved October 22, 2026) |
| Additional source | Gödel Machines, Sarvam's illusion of safety (retrieved October 23, 2026) |
| Additional source | Model metadata (retrieved October 17, 2026; October 28, 2026) |
| Additional source | OpenRouter: Popularity index (derived), CC BY 4.0 |
| Additional source | Vercel AI Gateway: Popularity index (derived), CC BY 4.0 |
| Additional source | Ollama: cumulative downloads for the popularity index |
| Additional source | Replicate: cumulative runs for the popularity index |
| Additional source | Google Trending Now: limited feed selection for Germany and the US |
| Additional source | Creative Commons Attribution 4.0 |