The model's architecture is extensively documented in official blog posts and Hugging Face model cards. It is a Mixture-of-Experts (MoE) transformer built from scratch using the NVIDIA NeMo framework and Megatron-LM. Specific technical details are provided, including the use of Multi-head Latent Attention (MLA) to compress the KV cache, a 32-layer depth (1 dense + 31 MoE), and a decoupled QK head dimension (q_head_dim=192, v_head_dim=128). The use of YaRN scaling for its 128K context window is also explicitly stated.

While the provider discloses the total token count (12 trillion) and the general categories of data (web, code, math, and multilingual content across 22 Indian languages), it lacks a detailed percentage breakdown of these sources. There is no public documentation on specific data filtering or cleaning methodologies, and no sample data or specific source names (e.g., specific web crawls or datasets) are provided beyond vague 'curated in-house' claims.

The tokenizer is publicly available via the Hugging Face repository. Technical documentation specifies its optimization for Indic languages, achieving fertility rates of 1.4 to 2.1 compared to the 4-8x seen in standard multilingual tokenizers. The vocabulary size and support for 22 Indian languages are clearly stated and verifiable through the provided configuration files.

The model provides exemplary transparency regarding its parameter density. It explicitly distinguishes between its 106B total parameters and its 10.3B active parameters per token. The MoE configuration is detailed, specifying 128 experts with a top-8 routing strategy and one shared expert. The architectural breakdown of the backbone (equivalent to a 10-13B dense transformer) is also publicly disclosed.

The provider identifies the hardware used (over 1,000 NVIDIA H100 GPUs at Yotta's Shakti cluster) and the framework (NVIDIA NeMo). However, it fails to disclose the total GPU hours, the duration of the training run, the estimated cost, or the carbon footprint associated with the training process.

While Sarvam provides scores for several standard benchmarks (MMLU, Math500, LiveCodeBench v6) and a custom Indic benchmark (IndiVibe), it does not provide the evaluation code or the specific prompts used for these results. Furthermore, the model has not yet appeared on major independent leaderboards like the Hugging Face Open LLM Leaderboard or LMSYS Chatbot Arena, making the self-reported results difficult to verify independently.

The model consistently identifies itself as Sarvam-105B and is transparent about its versioning and capabilities. It does not exhibit identity confusion or claim to be a model from a different provider. Documentation clearly outlines its intended use cases in agentic tasks and multilingual reasoning.

The model is released under the Apache 2.0 license, which is a standard, permissive open-source license. The terms are clear, allowing for both commercial and non-commercial use, modification, and distribution without conflicting proprietary restrictions.

Basic guidance on hardware requirements is available, such as the need for approximately 64GB of VRAM for non-quantized inference. The documentation mentions optimizations for NVIDIA Blackwell and the use of NVFP4 quantization, but detailed VRAM scaling for different context lengths and specific accuracy tradeoffs for various quantization levels (Q4, Q8) are not comprehensively documented for the end-user.

The model uses versioning (Sarvam-105B) and has a clear release date. However, there is no public changelog or established mechanism for tracking silent updates or performance drift. As a newly released model, a history of versioning and deprecation notices has not yet been established.