Krutrim: Multilingual Indic LLM Family

• Krutrim is a multilingual foundational LLM that addresses India's linguistic diversity by integrating the largest known Indic dataset and an optimized tokenizer.
• It employs a layered training pipeline with continual pre-training, instruction tuning, and DPO-based alignment to balance performance across Indic and English tasks.
• Krutrim variants power applications like speech translation and long-context literary QA, achieving competitive BLEU, ROUGE, and multilingual benchmark scores.

Searching arXiv for papers on Krutrim and closely related evaluations. Krutrim is a multilingual foundational LLM family designed for India’s linguistic landscape, with an explicit emphasis on Indic language coverage, balanced multilingual performance, and deployment in downstream systems that require robust English–Indic and Indic–English capabilities. In the foundational model paper, Krutrim is introduced as a 7B-parameter decoder-only Transformer trained on 2 trillion tokens, incorporating what the authors describe as the largest known Indic dataset, alongside an Indic-optimized tokenizer, instruction tuning, preference alignment via DPO, and integration with real-time web search for factual conversational use (Kallappa et al., 10 Feb 2025). Subsequent work positions Krutrim variants in distinct applied settings: Krutrim-1-instruct serves as the text-generation backbone in an end-to-end speech translation system for low-resource Indic speech translation (Wei et al., 25 Jul 2025), while Krutrim-2 12B is evaluated as an instruction-tuned multilingual LLM for long-context non-factoid question answering over Indic literary texts, where it emerges as the strongest model among the evaluated baselines (Khandelwal et al., 6 Jan 2026). At the same time, a human-centered audit of ableism detection shows that Krutrim 2 Instruct, despite being multilingual and Hindi-capable, remains insufficiently aligned with Indian disability perspectives in moderation-like settings, especially in Hindi (Phutane et al., 22 Jul 2025).

1. Foundational model and design objectives

Krutrim LLM is described as a multilingual foundational model built explicitly for India’s linguistic and cultural setting. The model is motivated by several challenges identified in the Indian context: extreme linguistic diversity; oral traditions and code-mixing; severe data sparsity and imbalance; tokenizer inefficiency on Indic scripts; and socio-economic and cultural diversity that complicates alignment and safety (Kallappa et al., 10 Feb 2025). The foundational paper states that Indic languages comprise approximately 1 percent of Common Crawl corpora even though India represents approximately 18 percent of the global population, and presents Krutrim as a direct response to this imbalance (Kallappa et al., 10 Feb 2025).

Architecturally, Krutrim is a causal decoder-only Transformer with approximately 7B parameters, 32 transformer layers, hidden dimension 4608, 48 attention heads, 8 KV heads using Grouped Query Attention, training-time context length 4096 tokens, ALiBi positional encoding, ReLU activation, and clipping of QKV matrix values (Kallappa et al., 10 Feb 2025). The combination of ALiBi and GQA is used to target long-context handling, reduced KV-cache size, and lower inference latency (Kallappa et al., 10 Feb 2025).

The tokenizer is trained from scratch using SentencePiece BPE on English plus Indic languages jointly, with the stated goal of reducing token-to-word ratios for Indic scripts and handling multiple scripts, morphological richness, compounding, and code-mixed text more effectively than generic open-source tokenizers (Kallappa et al., 10 Feb 2025). The paper does not provide an explicit vocabulary size, and no language tags or special token scheme are described.

A central design goal is balanced multilingual performance rather than optimization for English alone. The foundational paper states that Krutrim matches or exceeds models like LLaMA-2 on 10 out of 16 English tasks, with an average score of 0.57 versus 0.55, while also outperforming GPT-3.5 and Indic-specialized LLaMA derivatives on several Indic benchmarks (Kallappa et al., 10 Feb 2025). This positioning is important because it frames Krutrim not as a narrow regional model, but as a multilingual system that aims to preserve competitive English performance while substantially improving support for Indic languages.

2. Data curation, training procedure, and alignment

Krutrim is pre-trained on more than 2 trillion tokens drawn from open web and proprietary sources, including OpenWeb, a RedPajama subset, books, PubMed, Wikipedia, StackFast, and large curated Indic sources such as NDL (Kallappa et al., 10 Feb 2025). The paper emphasizes “hundreds of billions” of Indic tokens and describes the resulting corpus as the largest known distribution of Indic data to date (Kallappa et al., 10 Feb 2025). Data preparation includes deduplication, length filtering, quality filtering using approaches similar to Dolma, and additional Indic-specific cleaning to address script issues, encoding noise, and mixed or mislabeled text (Kallappa et al., 10 Feb 2025).

The pre-training objective is standard next-word prediction with cross-entropy loss, using NVIDIA H100 GPUs and approximately $10^{23}$ FLOPs (Kallappa et al., 10 Feb 2025). Checkpoints are evaluated every 20k steps and selected on the basis of validation metrics and loss curves (Kallappa et al., 10 Feb 2025). Beyond initial pre-training, the paper places strong emphasis on continual pre-training as a mechanism for adding new languages, domains, and capabilities while mitigating catastrophic forgetting. The described CPT mix includes approximately 25% original pre-training data and approximately 75% new high-quality data, followed by supervised fine-tuning on the same SFT data used in a PT-SFT baseline; CPT-SFT is reported to substantially outperform PT-SFT across seven task categories and multiple Indic languages in a Mean Opinion Score analysis (Kallappa et al., 10 Feb 2025).

Instruction tuning covers bidirectional translation, summarization, chain-of-thought reasoning, single- and multi-turn dialogue, safety around sensitive topics, general knowledge QA, coding and debugging, and chatbot self-identification and persona (Kallappa et al., 10 Feb 2025). The paper explicitly notes a trade-off between knowledge retention and instruction-following or creativity, cautioning that excessive SFT can overwrite pre-training knowledge (Kallappa et al., 10 Feb 2025).

For alignment, Krutrim uses Direct Preference Optimization rather than PPO-based RLHF. The reported DPO configuration uses learning rate $5 \times 10^{-7}$, $\beta = 0.1$, and approximately 20,000 preference instances focused initially on safety topics, with an explicit warning that an imbalanced language mix during DPO can induce forgetting or English-only drift (Kallappa et al., 10 Feb 2025). The model also undergoes an additional factuality-focused SFT stage after the initial instruction tuning. In the paper’s description, the instruction-tuned model initially showed approximately 33% hallucination on factual questions and approximately 14% hallucination or confirmation bias on adversarial or factually incorrect queries; the added SFT teaches the model to answer strictly from provided retrieved sources, detect ambiguity or factual incorrectness, and refrain when evidence is insufficient (Kallappa et al., 10 Feb 2025).

This training stack suggests that Krutrim should be understood not as a single frozen pre-trained checkpoint, but as a model family with a layered optimization pipeline: large-scale multilingual pre-training, continual pre-training for language and domain extension, supervised instruction tuning, preference alignment, and a factuality-specific adaptation stage.

3. Multilingual capabilities and benchmark profile

The foundational evaluation presents Krutrim as a model with strong and relatively balanced performance across multiple Indic languages. On IndicCOPA, using 3-shot prompts translated into Indic languages and evaluated with BERTScore, Krutrim scores 0.89 in Bengali, 0.83 in Gujarati, 0.86 in Hindi, 0.88 in Kannada, 0.88 in Malayalam, 0.87 in Marathi, 0.89 in Tamil, and 0.89 in Telugu, outperforming GPT-3.5 across all listed languages (Kallappa et al., 10 Feb 2025). On IndicQA, also evaluated with BERTScore in a zero-shot setting, Krutrim achieves 0.65 for Bengali, 0.64 for Gujarati, 0.64 for Hindi, 0.60 for Kannada, 0.66 for Malayalam, 0.58 for Marathi, 0.75 for Tamil, and 0.83 for Telugu, exceeding the cited results for Airavata, KAN-LLaMA, and TAM-LLaMA where comparisons are available (Kallappa et al., 10 Feb 2025).

On IndicSentiment, measured by accuracy in a 3-shot setup, Krutrim reports 0.95 for Bengali, 0.96 for Gujarati, 0.96 for Hindi, 0.95 for Kannada, 0.96 for Malayalam, 0.97 for Marathi, 0.94 for Tamil, and 0.95 for Telugu (Kallappa et al., 10 Feb 2025). On IndicTranslation, again using BERTScore, it scores 0.88 for Bengali, 0.89 for Gujarati, 0.95 for Hindi, 0.88 for Kannada, 0.89 for Malayalam, 0.92 for Marathi, and 0.88 for Telugu, with especially large margins over GPT-3.5 in non-Hindi languages (Kallappa et al., 10 Feb 2025). The paper also reports strong results on IndicXParaphrase, including 0.91 for Bengali, 0.97 for Hindi, 0.82 for Kannada, 0.90 for Malayalam, 0.94 for Marathi, and 0.61 for Telugu (Kallappa et al., 10 Feb 2025).

The same paper argues that Krutrim remains competitive in English. Across 16 English benchmarks including ARC, BIG-Bench, BoolQ, COPA, HellaSwag, Jeopardy, LAMBADA, LogiQA, MathQA, MMLU, OpenBookQA, PIQA, Simple Arithmetic, SQuAD, Winograd, and Winogrande, the model surpasses LLaMA-2 7B Chat on 10 tasks and achieves an average score of 0.569 versus 0.552 (Kallappa et al., 10 Feb 2025). Examples given include ARC 0.587 versus 0.517, BoolQ 0.854 versus 0.803, COPA 0.86 versus 0.78, LogiQA 0.378 versus 0.332, MMLU 0.51 versus 0.472, PIQA 0.78 versus 0.760, and SQuAD 0.408 versus 0.357 (Kallappa et al., 10 Feb 2025).

The paper further includes layer-wise probing on tasks such as GSM8K, calculation, Reclor, TruthfulQA, LAMA, and xMPS-Reason. The reported pattern is that factual knowledge is strongest in deeper layers, mathematical and logical reasoning improves toward upper layers, and cross-lingual math reasoning in English–Indic settings displays a similar layerwise structure to English-only reasoning, with the penultimate layer often outperforming the final layer (Kallappa et al., 10 Feb 2025). This suggests that the model’s cross-lingual competence is not confined to lexical transfer, but is represented throughout the network in a structured way.

A plausible implication is that Krutrim’s benchmark profile arises from three interacting factors explicitly described in the paper: a tokenizer optimized for Indic scripts, a much larger Indic token budget than generic multilingual models typically receive, and deliberate balancing of language distributions and context lengths during pre-training (Kallappa et al., 10 Feb 2025).

4. Krutrim in end-to-end speech translation

In HITSZ’s IWSLT 2025 Indic-track submission, Krutrim is used as the language-generation backbone in an end-to-end speech translation architecture derived from Dhwani and SALMONN (Wei et al., 25 Jul 2025). In this system, speech input passes through Whisper-large-v2 as a speech encoder, optionally through a BEATs encoder for non-speech audio, then through a Window-Level Query Transformer that maps encoder outputs into a token sequence compatible with the LLM, after which Krutrim-1-instruct generates the target-language text (Wei et al., 25 Jul 2025). The architecture uses Krutrim as the decoder or generator rather than as a post-editor or reranker.

Training is parameter-efficient. For English-to-Indic, WhisperSE is frozen; for Indic-to-English, WhisperSE is trainable for the first epoch and then frozen; BEATs remains frozen; the Q-Former is fully trainable; and Krutrim is adapted using LoRA rather than full fine-tuning, with rank $r = 8$ and $\alpha = 32$ (Wei et al., 25 Jul 2025). The paper provides the standard LoRA formulation

$W' = W + \Delta W,\quad \Delta W = BA,$

with low-rank factors $A$ and $B$, and describes Krutrim’s ST training in terms of standard token-level cross-entropy over target sequences conditioned on the Q-Former outputs (Wei et al., 25 Jul 2025).

The speech data come exclusively from the IWSLT Indic track. For English-to-Indic, each of Bengali, Hindi, and Tamil includes 680.9 hours of training audio, 40.8 hours of development audio, and 93.2 hours of test audio, totaling 814.9 hours per pair (Wei et al., 25 Jul 2025). For Indic-to-English, the totals are 160.3 hours for Bengali-to-English, 656.2 hours for Hindi-to-English, and 481.4 hours for Tamil-to-English (Wei et al., 25 Jul 2025). Each datum includes an audio clip, a transcription, and a translation. For English-to-Indic training, examples are split into short sequences with English transcriptions under 400 characters and long sequences with transcriptions of at least 400 characters, enabling two-stage fine-tuning from short to long clips (Wei et al., 25 Jul 2025).

The reported test BLEU scores show average performance of 28.88 for English-to-Indic and 27.86 for Indic-to-English (Wei et al., 25 Jul 2025). Per-direction English-to-Indic BLEU scores are 27.00 for English-to-Bengali, 33.84 for English-to-Hindi, and 22.81 for English-to-Tamil (Wei et al., 25 Jul 2025). Indic-to-English BLEU scores are 25.02 for Bengali-to-English, 39.29 for Hindi-to-English, and 19.27 for Tamil-to-English (Wei et al., 25 Jul 2025). The paper explicitly notes that all reported systems use Krutrim as the LLM component, so every BLEU value reflects the integrated Whisper–Q-Former–Krutrim stack (Wei et al., 25 Jul 2025).

The same work investigates a Chain-of-Thought formulation for Indic-to-English only. Here, Krutrim is fine-tuned to output both an Indic transcription and an English translation in a structured multi-segment format, corresponding to a “first transcribe, then translate” strategy (Wei et al., 25 Jul 2025). On the development set, the average CoT parsing success rate is 66.54%, with 68.18% for Bengali-to-English, 71.00% for Hindi-to-English, and 60.43% for Tamil-to-English (Wei et al., 25 Jul 2025). On the parsable subset, BLEU reaches 28.13 for Bengali-to-English with $\Delta = +2.592$, 38.49 for Hindi-to-English with $\Delta = +6.780$, and 34.77 for Tamil-to-English with $5 \times 10^{-7}$0 (Wei et al., 25 Jul 2025). The paper identifies the central failure mode as inconsistent adherence to the required CoT output format, including unclear segmentation, additional commentary, or misordered segments, and frames future work as improving instruction-following and format adherence for Indic LLMs in multimodal ST (Wei et al., 25 Jul 2025).

This applied study is important because it demonstrates that Krutrim can function as the decoder in a multimodal sequence-to-sequence pipeline under low-resource conditions, but it also reveals a specific limitation: instruction tuning sufficient for ordinary generation may still be inadequate when multimodal structured output must be strictly parseable.

5. Krutrim-2 in long-context literary question answering

The LittiChoQA study evaluates Krutrim-2 12B as one of six instruction-tuned multilingual LLMs on long-context, non-factoid question answering over literary texts in 17 Indic languages (Khandelwal et al., 6 Jan 2026). Krutrim-2 is identified as a CausalLM with more than 1B parameters and used via the instruction-tuned checkpoint krutrim-ai-labs/Krutrim-2-instruct (Khandelwal et al., 6 Jan 2026). Within the paper’s evaluated group, it is the largest model at 12B parameters and is reported to be the strongest model across all tested settings (Khandelwal et al., 6 Jan 2026).

LittiChoQA contains over 270K QA pairs drawn from naturally authored literary texts, including stories, folk literature, books, and novels collected from the open web (Khandelwal et al., 6 Jan 2026). Although the dataset contains both factoid and non-factoid items, model fine-tuning and evaluation are restricted to non-factoid question-answer pairs, with a train/dev/test split of 70:10:20 (Khandelwal et al., 6 Jan 2026). Krutrim-2 is evaluated in three regimes: full context, in which the model receives the full story or story segment; Answer Paragraph Selection, in which a fine-tuned APS model selects the most relevant paragraph(s); and vector-based retrieval, in which paraphrase-multilingual-MiniLM-L12-v2 embeddings retrieve top paragraphs (Khandelwal et al., 6 Jan 2026).

Fine-tuning is done with PEFT and LoRA using AdamW, learning rate $5 \times 10^{-7}$1, 3 epochs, LoRA parameters $5 \times 10^{-7}$2 and $5 \times 10^{-7}$3, 4-bit NF4 quantization, bf16 training precision, and two NVIDIA A100 GPUs (Khandelwal et al., 6 Jan 2026). The paper does not specify a special long-context architecture or explicit maximum sequence length; long-context operation is instead implicit in providing full stories within available GPU limits.

Krutrim-2 attains the best reported results in every context regime. With full context, it achieves ROUGE-1 37.8, ROUGE-2 16.7, ROUGE-L 27.1, and STS MuTe 76.1 (Khandelwal et al., 6 Jan 2026). With APS-shortened context, the scores are ROUGE-1 35.2, ROUGE-2 15.2, ROUGE-L 25.2, and STS MuTe 74.9 (Khandelwal et al., 6 Jan 2026). With vector-retrieval context, the scores are ROUGE-1 30.5, ROUGE-2 12.2, ROUGE-L 21.8, and STS MuTe 71.4 (Khandelwal et al., 6 Jan 2026). Across all three settings, these scores exceed those of OpenHathi, Sarvam-1, Qwen2.5, Aya, and Llama 3.1 (Khandelwal et al., 6 Jan 2026).

The study emphasizes the trade-off between answer quality and efficiency. Full-context models can answer only about 900 questions within the fixed compute budget, whereas shortened-context models can answer more than 20K questions (Khandelwal et al., 6 Jan 2026). For Krutrim-2 specifically, the shift from full context to APS reduces STS MuTe only from 76.1 to 74.9, while substantially increasing throughput (Khandelwal et al., 6 Jan 2026). The authors therefore present APS as the most effective context-shortening method relative to naive vector retrieval.

Qualitative analysis further supports Krutrim-2’s status in this benchmark. Human annotators rank the reference answer first and Krutrim-2 second on three sampled full-context instances, and under shortened-context settings Krutrim-2 remains consistently among the top three (Khandelwal et al., 6 Jan 2026). A complementary GPT-4.1-based evaluation reports that in shortened-context settings Krutrim-2 is ranked first in 12 of 17 languages (Khandelwal et al., 6 Jan 2026). The paper does not provide a fine-grained error taxonomy, but the reported degradation under vector retrieval is interpreted as evidence that answer quality remains sensitive to missing or incomplete context (Khandelwal et al., 6 Jan 2026).

6. Societal evaluation, limitations, and contested performance

A distinct line of evidence comes from the audit of ableism detection across Western and Indic LLMs, where Krutrim 2 Instruct is evaluated alongside GPT-4o, Gemini 2.0 Flash, Claude 3.7 Sonnet, Llama 3.1 70B, Nanda 10B Chat, Gajendra v.01, and Airavata (Phutane et al., 22 Jul 2025). The task uses a dataset of 300 comments: 100 English comments, 100 Hindi-formal translations, and 100 Hindi-casual translations, all rated on 0–10 toxicity and ableism scales by both models and people with disabilities (Phutane et al., 22 Jul 2025). Human evaluation involves 175 PwD total, including 130 in the U.S. and 45 in India, with nonparametric analyses based on Wilcoxon signed-rank tests, Kruskal–Wallis tests, and Spearman correlations (Phutane et al., 22 Jul 2025).

For English comments, Krutrim’s mean toxicity score on ableist comments is 5.96 with SD 2.13, and its mean ableism score is 6.37 with SD 2.35 (Phutane et al., 22 Jul 2025). On non-ableist comments, the corresponding means are 1.63 for toxicity and 0.13 for ableism (Phutane et al., 22 Jul 2025). Indian PwD, by contrast, report 7.13 toxicity and 6.97 ableism on ableist comments, and 2.64 toxicity and 2.43 ableism on non-ableist comments (Phutane et al., 22 Jul 2025). The paper characterizes the broader pattern as one in which Western LLMs overestimate ableist harm, while Indic LLMs, including Krutrim, underestimate it (Phutane et al., 22 Jul 2025).

The language-sensitivity findings are especially notable. For Krutrim, toxicity differs significantly between English and Hindi-formal with $5 \times 10^{-7}$4, and ableism differs significantly between English and Hindi-casual with $5 \times 10^{-7}$5 and between English and Hindi-formal with $5 \times 10^{-7}$6 (Phutane et al., 22 Jul 2025). These results indicate that Krutrim rates the same underlying content as more harmful in English than in Hindi, whereas Indian PwD show no significant language-based differences in harm perception (Phutane et al., 22 Jul 2025). The study therefore argues that Krutrim is multilingual but not adequately multicultural in this domain (Phutane et al., 22 Jul 2025).

The qualitative error patterns are also specific. The paper reports that Krutrim sometimes misses microaggressions, supplies incorrect causal explanations for harm, and misreads relational or register cues in Hindi (Phutane et al., 22 Jul 2025). One example concerns the comment “Are you sure you’re not just doing this for attention?”, where Krutrim explains the harm as implying that disabled people are not capable of experiencing genuine emotions, rather than recognizing the more direct accusation that the person is faking or exaggerating disability (Phutane et al., 22 Jul 2025). Another example concerns the comment “Your prosthetic leg is so hot,” where demographic prompting that specifies Indian context leads Krutrim to reinterpret the statement as a compliment rather than a derogatory or discriminatory statement (Phutane et al., 22 Jul 2025). The paper concludes that Krutrim should not be used as a sole arbiter of ableist harm, especially in Hindi, and recommends local disability-centered datasets, Indian PwD involvement in model design and evaluation, and language-aware human-in-the-loop moderation (Phutane et al., 22 Jul 2025).

This line of evidence complicates a simple benchmark-based view of model quality. Krutrim performs strongly on formal NLP tasks such as translation, reading comprehension, and long-context QA, yet the audit indicates that those strengths do not automatically translate into reliable culturally grounded harm recognition. A plausible implication is that multilingual competence and local-language coverage are necessary but not sufficient for sociocultural alignment.

7. Interpretation and trajectory of the Krutrim model family

Across the cited studies, Krutrim appears less as a single static model than as a family of Indic-oriented LLMs occupying multiple positions in the current landscape of multilingual AI. The foundational 7B model is presented as a decoder-only Transformer optimized for Indic linguistic coverage, with strong benchmark performance, deliberate data balancing, an Indic-specific tokenizer, and a WebRAG deployment path for factual conversational applications (Kallappa et al., 10 Feb 2025). Krutrim-1-instruct is then used as a generator in a multimodal speech translation stack, where LoRA adaptation and a Q-Former bridge enable end-to-end speech-to-text translation for English, Hindi, Bengali, and Tamil (Wei et al., 25 Jul 2025). Krutrim-2 12B appears as a larger instruction-tuned multilingual model that leads a competitive set of baselines on long-context literary QA in 17 Indic languages (Khandelwal et al., 6 Jan 2026). In contrast, Krutrim 2 Instruct shows substantial shortcomings in culturally grounded moderation-like evaluation, particularly in Hindi ableism detection and explanation (Phutane et al., 22 Jul 2025).

Several themes recur across these settings. First, Krutrim’s strongest evidence lies in tasks where rich multilingual pre-training, Indic tokenization, and instruction tuning directly support generation or comprehension. This is visible in the foundational benchmark tables, in the BLEU scores obtained in speech translation, and in the STS MuTe and ROUGE metrics for literary QA (Kallappa et al., 10 Feb 2025, Wei et al., 25 Jul 2025, Khandelwal et al., 6 Jan 2026). Second, parameter-efficient adaptation is central to its applied use. Both the speech translation system and the LittiChoQA experiments rely on LoRA rather than full fine-tuning, with ranks of $5 \times 10^{-7}$7 in both cases, indicating that Krutrim is already being treated as a reusable multilingual backbone rather than a model that must be retrained wholesale for each task (Wei et al., 25 Jul 2025, Khandelwal et al., 6 Jan 2026). Third, instruction-following quality is a recurrent bottleneck. In speech translation, CoT gains are large on the parsable subset but are limited by failure to maintain the required structured format (Wei et al., 25 Jul 2025). In ableism detection, explanations are often thin, generic, or misaligned with local human reasoning (Phutane et al., 22 Jul 2025).

The future directions named in the papers are correspondingly concrete. The foundational paper emphasizes expanding language coverage via continual pre-training, adding richer domain knowledge, improving alignment, and continuing to improve WebRAG and factual reliability (Kallappa et al., 10 Feb 2025). The speech translation paper explicitly identifies improved instruction-following and improved Indic generation quality as next steps, especially to make CoT-based ST more reliable and to strengthen English-to-Indic performance (Wei et al., 25 Jul 2025). The LittiChoQA paper points toward dynamic context filtering, integrated long-context models, and further exploration of multilingual transfer and cross-lingual generalization (Khandelwal et al., 6 Jan 2026). The ableism audit recommends re-centering local disability experience, extending evaluation to more Indic languages, and retraining or fine-tuning on data annotated by Indian PwD with explicit attention to intersectionality and register (Phutane et al., 22 Jul 2025).

Taken together, these studies depict Krutrim as a technically strong and distinctly Indic-oriented LLM family whose empirical record is bifurcated. On conventional multilingual NLP and generative benchmarks, Krutrim is repeatedly competitive or best-in-group. On tasks requiring strict structured adherence or culturally situated normative judgment, the evidence is more cautionary. This suggests that the family’s long-term significance will depend not only on scaling and benchmark gains, but also on whether future iterations can convert multilingual breadth into more reliable multimodal control, context-sensitive instruction following, and deeper sociocultural alignment.