Assess · Sarvam-30B, Sarvam-105B and Sarvam-105B-1711

Can you own it?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input D ยท 44.4/100

This page is a projection of the one entry record, the Use & modify and Transparency factors that Assess covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Intended & out-of-scope use

The cards describe reasoning, coding, agentic and Indian-language use for the 105B, and low-latency conversation and voice-agent tool calls for the 30B (ev-card-105b, ev-card-30b). They state no out-of-scope uses and no intended-purpose statement.

The hosted Terms prohibit several uses, including professional advice without review. Those Terms govern Sarvam's service, and the weights carry no acceptable-use policy (ev-api-docs, ev-tos).

Known limitations, bias & failure modes

No card lists a limitation, bias or risk (ev-card-105b, ev-card-30b, ev-card-1711).

Artificial Analysis reports high hallucination rates for both March builds, with AA-Omniscience indexes of -60 (105B) and -72 (30B). It also reports TerminalBench Hard scores of 1.5% and 2.3% (ev-aa).

The API documentation lists one limitation. Reasoning is on by default and its tokens count against max_tokens (ev-api-docs). Open Hugging Face threads report an import error with newer Transformers and thinking that cannot be turned off (ev-discussions).

Openness tier & components

The tier is open_weights.

ComponentRatingEvidence
WeightsopenUngated safetensors for all builds, plus official FP8 and GGUF (ev-hub, ev-quant-code)
Training dataclosedThe blog describes categories only. No dataset is released (ev-blog)
Training codeclosedThe GitHub organisation holds forks and SDKs, no training code (ev-quant-code)
EvaluationpartialThe indivibe prompt set is public under CC-BY-4.0. Results are charts, with no judge outputs or code (ev-indivibe)
DocumentationpartialCards and a blog exist. There is no technical report (ev-no-report)
LicenceopenApache-2.0 declared, with no LICENSE file (ev-license)

Only post-trained reasoning models are released. No base checkpoint repo exists in the organisation (ev-hub). A community comment claiming base and instruct variants is wrong against the repo list (ev-discussions).

License terms & permitted use

Apache-2.0 is declared in the front matter of every card and in the sentence "open-sourced under the Apache License" (ev-license). No weights repo holds a LICENSE file. The cards give no version number and no link. The standard text grants copyright and patent licences and allows commercial use.

Sarvam's own legal documents add a layer the weights repos never mention. Terms of Service v2.0 clause 10.5 bars using the Offerings or their output to train AI systems and bars discovering model weights. The EULA defines Software to include "proprietary language models", bars derivative works and bars training competing models (ev-tos, ev-eula).

Neither document mentions the Apache-2.0 weights or carves them out. The text does not say whether they reach a copy downloaded under Apache-2.0. Get legal review before building a derivative.

Supply-chain & provenance

The canonical source is the sarvamai Hugging Face organisation. It is not verified. Weights are safetensors with LFS sha256 values, no pickle files, and a completed hub scan with no flagged files on the main repos. The GGUF repos show no completed scan (ev-hub).

Nothing is signed. Loading needs trust_remote_code=True, and the 30B and 105B repos ship hotpatch_vllm.py, which installs vLLM 0.15.0 and edits its registry (ev-card-105b, ev-hub).

Sarvam says the models were trained from scratch on in-house data (ev-blog). That is a publisher claim. Artificial Analysis repeats it without checking (ev-aa). No base model is named for the 30B or the 105B. The 1711 build declares sarvamai/sarvam-105b as its base (ev-card-1711).

EU AI Act posture

These are general-purpose AI models. At about 7.4e23 FLOPs for the 105B and 2.3e23 for the 30B (OneHill estimates from publisher token counts and active parameters), they are far below the 1e25 systemic-risk line.

Sarvam publishes no technical documentation package, no training-data summary in the AI Office template and no copyright policy (ev-no-report, ev-blog). Its Code of Practice status is unknown.

The open-source exemption is uncertain. Apache-2.0 is declared, and Sarvam also charges for API access to the same models (ev-token-factory). A deployer assembles Annex XI and XII style information alone.

Benchmarks & evaluation

All card and blog numbers are the publisher's own. The 105B card reports 98.6 on Math500, 78.7 on GPQA Diamond and 68.3 on tau2-Bench. The 30B card reports 97.0, 66.5 and 45.7. The 1711 card reports 80.4 on GPQA Diamond, 85.5 on tau2-Bench and 79.1 on an internal Indic Suite (ev-card-105b, ev-card-30b, ev-card-1711).

The blog and the 30B card disagree on AIME 25: 80.0 against 88.3 (ev-blog, ev-card-30b).

Artificial Analysis scored the 105B at 18 and the 30B at 12 on its Intelligence Index in April 2026. Both trail every peer it lists. The score covers the March builds only (ev-aa).

The Indian-language claim rests on a pairwise LLM-judge comparison over the public indivibe prompts. The blog reports average win rates of 90% (105B) and 89% (30B) against models it does not list in text. Per-language results are charts (ev-blog, ev-indivibe).

Independent safety evaluation

No safety, red-team or jailbreak evaluation of any build was found, from Sarvam or from a third party. The blog describes safety fine-tuning data with India-specific risks and red-team prompts and gives no results (ev-blog).

How this scores

The ownership factors this domain covers, drawn from the one entry record.

1

Use and modify freelyCan you run, modify and adapt it with no gate and no field-of-use trap?

Moderate

Apache-2.0 is declared, the weights are ungated and fine-tuning is possible. The licence is a card declaration with no LICENSE file, and Sarvam's Terms and EULA add limits on derivative works and training that do not mention the weights (ev-license, ev-tos, ev-eula).

How this scores (AOI sub-dimensions)
Openness3/5how much is released - weights, data, code, licence - and how freelyWeights are open under an Apache-2.0 declaration, with cards, a blog and a public 22-language evaluation prompt set.
Legal2/5how permissive and clean the licence is for real commercial useApache-2.0 is declared but no repo holds a LICENSE file, and Sarvam's Terms and EULA add limits on derivative works and training that do not mention the weights.
2

TransparencyDo you know what it is: weights, training, behaviour, and legible terms?

Moderate

Weights, cards and a blog are public. Training data, training code, post-training recipes, base checkpoints and a technical report are not (ev-card-105b, ev-blog, ev-no-report).

How this scores (AOI sub-dimensions)
Provenance2/5how well we can trace and verify what went into the modelThe canonical source is the sarvamai Hugging Face organisation, which is not verified.
Governance3/5how accountable and well-documented the publisher isSarvam AI is an accountable company in Bengaluru with a release record: the March release, fix pull requests in September and November, and the 1711 build in November.
What this means for adoptionYou can run and fine-tune Apache-2.0-declared weights, but you get no data, code, recipes or safety case, and the licence is a card declaration with no LICENSE file. Sarvam's own Terms and EULA restrict derivative works and training without mentioning the weights. Ownership rises if Sarvam adds a LICENSE file, clarifies those documents and publishes a safety evaluation.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Model cardread2026-11-19
The 105B card declares Apache-2.0 and lists 22 language codes plus English.
Model cardread2026-11-19
The 30B card declares Apache-2.0 and the same language codes.
Model cardread2026-11-19
The 105B-1711 cards declare Apache-2.0 and say the model shares the Sarvam-105B base and is further post-trained for tool calling, instruction following, agentic tasks and Indian languages.
Licenceread2026-11-19
All eight weights repos declare Apache-2.0 in the card front matter and none holds a LICENSE file (raw LICENSE returns HTTP 404).
Documentationread2026-11-19
The sarvamai Hugging Face organisation is not verified and lists 17 ungated model repos.
Vendor announcementread2026-11-19
The launch blog of 6 March 2026 says both models were trained from scratch on in-house data, 16T tokens for the 30B and 12T for the 105B, with the 10 most-spoken Indian languages emphasised in pre-training (publisher claims).
Documentationread2026-11-19
The indivibe dataset holds 4,840 evaluation prompts in the 22 scheduled languages, including Bodo, in native and romanised script, under CC-BY-4.0.
Documentationread2026-11-19
Sarvam's API documentation lists 10 Indian languages plus English for Sarvam-105B and for the deprecated Sarvam-30B.
Documentationread2026-11-19
The platform documentation says workspace data is retained indefinitely until an owner sets a period.
Terms of serviceread2026-11-19
Terms of Service v2.0, effective 29 July 2026, clause 10.5 bars using the Offerings or their output to train or fine-tune AI systems and bars discovering model weights.
Privacy Policyread2026-11-19
The Privacy Policy of 29 July 2026 headlines model training as opt-in but says Sarvam trains on inputs and outputs unless the user opts out.
Terms of serviceread2026-11-19
The EULA of 22 May 2026 defines Software to include 'proprietary language models' and bars derivative works.
Securityread2026-11-19
The Trust Center claims ISO 27001, SOC 2 Type II, India-only residency and no use of one customer's data to train models for other customers.
Third-party analysisread2026-11-19
Artificial Analysis scored Sarvam 105B at 18 and Sarvam 30B at 12 on its Intelligence Index in April 2026.
Documentationread2026-11-19
Official FP8 builds of both models and Q4_K_M GGUF builds (19.6 GB for the 30B, 64.2 GB for the 105B) exist.
Documentationread2026-11-19
Community threads ask for the training data and for newer checkpoints and have no Sarvam reply.
Vendor announcementread2026-11-19
The Token Factory post of 16 November 2026 says Sarvam serves the 105B in MXFP8 with DSpark speculative decoding from GPU clusters in India.
Technical_reportunverified2026-11-19
No technical report, paper or data statement for these models was found.
Documentationunverified2026-11-19
The AIKosh page for the 105B, linked from the blog as a second download source, is a JavaScript app and returns no text.