sarvam-30b
Sarvam-30B is a 30-billion-parameter open-source Mixture-of-Experts (MoE) language model optimized for multilingual support across 22 Indian languages and efficient deployment in resource-constrained environments. With only 2.4B active parameters, it balances reasoning, coding, and conversational capability while maintaining lower memory footprint than dense models of similar scale. Licensed under Apache 2.0 with no gating restrictions, it is suitable for self-hosted and custom LLM applications.
Key facts
Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.
| Field | Value |
|---|---|
| Developer | sarvamai |
| Parameters | 32.2B |
| Context window | Unknown |
| License | apache-2.0 — OSI-approved |
| Modality / task | text-generation |
| Gated on HuggingFace | No |
| Downloads | 55.3k |
| Likes | 211 |
| Last updated | 2026-03-23 |
| Source | sarvamai/sarvam-30b |
What sarvam-30b is
Sarvam-30B uses a 19-layer MoE architecture with 128 routed experts plus one shared expert, top-6 routing, grouped KV attention (4 heads), and dense FFN intermediate size 8192 with MoE intermediate size 1024. It employs high rope_theta (8e6) for long-context stability without RoPE scaling, auxiliary-loss-free router balancing, and 2.5x routed scaling factor. The model supports context lengths up to 65,536 tokens in benchmarking. Evaluated with transformers, safetensors, and custom_code=true, requiring trust_remote_code. Inference backends include HuggingFace, vLLM (via hot-patch or source fork), and SGLang.
Run sarvam-30b locally
Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.
from transformers import pipelinepipe = pipeline("text-generation", model="sarvamai/sarvam-30b")out = pipe("Explain retrieval-augmented generation in one sentence.", max_new_tokens=128)print(out[0]["generated_text"])Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.
How you'd run it
A typical self-hosted path — open weights, an inference server, your application.
DEV.co builds each layer — from GPU infrastructure to the application.
Best use cases
Running & fine-tuning it
ESTIMATE (requires independent verification): 30B parameters with MoE architecture and 2.4B active parameters. In bfloat16 precision (~60 GB disk storage), activation-only inference approximately 5–12 GB VRAM per GPU depending on batch size and routing sparsity. Multi-GPU inference with tensor parallelism (tp_size=2–8) recommended for throughput. vLLM example uses tp_size=8. Quantized variants (int8, int4) would reduce footprint further but not mentioned in card.
Model card does not explicitly document LoRA, QLoRA, or full fine-tuning support. Custom code required (trust_remote_code=True). Standard HuggingFace transformer-based LoRA should be feasible given transformers framework compatibility, but adapter compatibility and training stability on MoE architecture not documented. Recommend testing on small dataset before production fine-tuning.
When to avoid it — and what to weigh
- Strict Long-Context Requirements Beyond 65K Tokens — Benchmark evaluation maxes at 65,536 tokens. Production context length limits are not stated. If applications require 100K+ token context windows with proven performance, requires independent testing.
- Minimal Agentic Performance Needed — Agentic benchmarks show mixed results: BrowseComp (35.5%), SWE Bench (34.0%), τ² Bench (45.7%) trail some competing models (e.g., GLM 4.7 Flash at 79.5% τ²). For mission-critical agent systems, compare against specialized alternatives.
- Maximum Inference Latency Sensitivity — MoE routing overhead and sparse expert selection may introduce variable latency. Inference speed benchmarks not provided. Dense alternatives may offer more predictable throughput for real-time applications.
- Weak Coverage of Non-Indian Language Domains — Model is optimized for Indian languages and English. Performance on other non-English languages not clearly documented. Not recommended as primary choice for Japanese, Chinese, or Arabic-heavy workloads.
License & commercial use
Apache License 2.0 (SPDX: apache-2.0). Permissive OSI-approved open-source license. No gating or model agreement required (gated=false). Permits commercial use, modification, and distribution with attribution and Apache 2.0 boilerplate.
Apache 2.0 is a permissive open-source license that explicitly permits commercial use and distribution. No royalty obligations, proprietary license fees, or commercial restrictions apply. Users may deploy, modify, and resell applications incorporating Sarvam-30B. However, liability disclaimers and patent grant limitations in Apache 2.0 should be reviewed by legal counsel for high-risk deployments. No warranty or indemnification provided by sarvamai.
DEV.co evaluation signals
Editorial assessment — not user reviews. Directional, with an explicit confidence level.
| Signal | Assessment |
|---|---|
| Maintenance | Active |
| Documentation | Adequate |
| License clarity | Clear |
| Deployment complexity | Moderate |
| DEV.co fit | Strong |
| Assessment confidence | High |
Model requires trust_remote_code=True for HuggingFace and vLLM loading, introducing code execution risk if model repository is compromised or tampered. Standard LLM security considerations apply: no built-in guardrails or safety filters documented; evaluate output for bias, toxicity, and hallucinations before production deployment, especially in sensitive Indian language contexts. No security audit or adversarial robustness testing mentioned. No rate-limiting, input sanitization, or prompt injection mitigations documented.
Alternatives to consider
Nemotron-3-Nano-30B (NVIDIA)
Similar 30B parameter class, stronger agentic benchmarks (SWE Bench 38.8%, τ² Bench 49.0%), comparable coding performance. Closed-source but NVIDIA-backed. Choose if agentic capability is priority over Indian language support.
Qwen3-30B-Thinking (Alibaba)
30B dense model with explicit reasoning/thinking mode, superior MMLU Pro (80.9%) and math (Math500 97.6%), strong agentic benchmarks (τ² 47.7%). Better for pure reasoning workloads; MoE trade-off may not be necessary.
Mistral 3.2-24B
Smaller dense model, permissive license, strong HumanEval (92.9%), well-documented. Lower resource footprint but weaker knowledge benchmarks (MMLU 80.5%, MMLU Pro 69.1%) and no Indian language specialization.
Ship sarvam-30b with senior software developers
Sarvam-30B offers permissive Apache 2.0 licensing and strong Indian language support for self-hosted applications. Evaluate memory and latency requirements on your hardware, test fine-tuning if needed, and integrate via HuggingFace, vLLM, or SGLang. Contact Devco to architect your deployment.
Talk to DEV.coRelated open-source tools
Surfaced by semantic similarity across the DEV.co open-source index.
Related on DEV.co
Explore the category and the services that help you build with it.
sarvam-30b FAQ
Can I use Sarvam-30B in a commercial product without paying licensing fees?
What GPU memory is needed to run Sarvam-30B inference?
Can I fine-tune Sarvam-30B on my own data?
Does Sarvam-30B support languages outside India?
Work with a software development agency
DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If sarvam-30b is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.
Ready to Deploy Multilingual AI in Your Infrastructure?
Sarvam-30B offers permissive Apache 2.0 licensing and strong Indian language support for self-hosted applications. Evaluate memory and latency requirements on your hardware, test fine-tuning if needed, and integrate via HuggingFace, vLLM, or SGLang. Contact Devco to architect your deployment.