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

Can you run it?

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

This page is a projection of the one entry record, the Reliability and Data control factors that Implement 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

Install & run

Use Transformers with trust_remote_code=True, SGLang or vLLM (ev-card-105b). The repos ship configuration_sarvam_moe.py and modeling_sarvam_moe.py, which run on your machine.

For the March builds, vLLM needs the Sarvam fork or hotpatch_vllm.py. The script installs vLLM 0.15.0, adds two registry entries and downloads model executors. Review it first.

The 1711 cards give vllm serve commands for vLLM 0.33.0 with --trust-remote-code, --tool-call-parser glm45 and --reasoning-parser glm45 (ev-card-1711). An optional dspark/ draft model enables speculative decoding.

Pin the revision with revision="<commit-sha>". The weights are ungated and need no login (ev-hub).

Hardware & VRAM requirements

No minimum is stated for the March builds. File sizes give a floor (ev-hub, ev-quant-code, ev-card-1711).

BuildSize on disk
sarvam-105b (float32)424.2 GB
sarvam-105b-fp8112.6 GB
sarvam-105b-gguf (Q4_K_M)64.2 GB
sarvam-105b-1711-bf16218.3 GB
sarvam-105b-1711 (NVFP4, MXFP8)70.2 GB
sarvam-30b (float32)128.6 GB
sarvam-30b-fp838.7 GB
sarvam-30b-gguf (Q4_K_M)19.6 GB

The card examples use tensor parallelism of 4 (SGLang) to 8 (vLLM) for the 105B and 2 for the 30B. The 1711 card reports throughput on one B200 (128K context) and on one 96 GB RTX PRO 6000 Blackwell (50K context).

Serving stacks

Transformers, SGLang and vLLM are documented. vLLM needs the fork or hot-patch for the March builds, and a native pull request was open at release (ev-card-105b). The 1711 cards test vLLM 0.33.0 (ev-card-1711). The GGUF cards give llama.cpp build and run commands (ev-quant-code).

Ollama and MLX are undocumented. A request for MLX support is open (ev-discussions).

Safe-deployment controls & Deployment Ceiling

Deployment Ceiling T1, conditional. Use it for internal tasks with a human reading every output.

Before any step up, add your own guard classifier, safety and jailbreak tests, a reviewed copy of the custom Python files and sandboxed tool execution. Sarvam publishes no safety evaluation and no guard model (ev-blog, ev-card-1711).

Available quantizations

Sarvam publishes FP8 builds of both models, Q4_K_M GGUF builds of both, and a mixed NVFP4 and MXFP8 build of 105B-1711. The FP8 cards say the weights were quantised with NVIDIA ModelOpt (ev-quant-code, ev-hub).

Fine-tuning & adaptation

Apache-2.0 weights allow fine-tuning, and Prism is a Sarvam LoRA fine-tune of the 105B (ev-token-factory). A merged staff pull request restores differentiable single-rank MoE dispatch for training and fails on expert-parallel training (ev-discussions).

No recipe, dataset or post-training code is published. Check the Terms and EULA position on derivatives before you ship one (ev-tos, ev-eula).

API / OpenAI-compatible integration

Self-hosted vLLM exposes an OpenAI-compatible /v1/chat/completions endpoint. Thinking is on by default and is turned off with chat_template_kwargs (ev-card-1711).

Sarvam's hosted API is also OpenAI-compatible and carries its own terms, covered in the support page (ev-api-docs).

How this scores

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

3

ReliabilityIs it reliable and good enough for the job?

Weak

Safety scores 1: no safety evaluation, no guard model and silent 1711 cards. Operational scores 3 (ev-blog, ev-card-1711, ev-hub).

How this scores (AOI sub-dimensions)
Operational3/5how practical it is to run, serve and maintain in productionThe models run on Transformers with trust_remote_code, SGLang and vLLM.
Safety1/5whether misuse risks are evaluated and guardrails are providedThe blog says the fine-tuning data covered India-specific risks and red-team prompts.
4

Doesn't extract your dataDoes running it keep your knowledge and data yours?

Moderate

Weights run on your own hardware. The custom Python and the vLLM hot-patch script need review, and the hosted API terms conflict on training and retention (ev-hub, ev-privacy, ev-eula).

How this scores
Not a scored AOI dimension. For a self-hosted model, data-control is a structural property of running the weights yourself, strong by default unless the model phones home or the licence claws back rights. For a hosted API this factor is the retention + train-on-inputs + residency read, scored from the binding terms.
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.