Gnani AI marks the next frontier in India’s race for Sovereign AI

The Problem

India pays a Language Tax on every AI request. Evon v3.3 is built to remove it.

The Model

A hybrid mixture-of-experts model, sized for one node

Head to Head · India

Evon v3.3 beats a model 3.5× its size on Indian languages

Sarvam-105B carries 105 billion parameters against Evon v3.3's 30 billion. On MILU — 79,608 items across 11 Indian languages and 41 subjects, identical items and prompts for every model — Evon v3.3 leads Sarvam-105B on 10 of 11 languages and scale-matched Sarvam-30B on all 11.

MILU accuracy, language by language

Capability categories — Indian-language benchmarks vs Sarvam

Grouping the Indic benchmark suite by capability shows where the advantage sits. Evon v3.3 leads Sarvam-30B on every category reported here. Against Sarvam-105B it leads on instruction following, sits level on question answering and translation — both gaps inside a fifth of a point — and trails on general knowledge and reasoning.

Head to Head · Global

How Evon v3.3 compares with OpenAI, Google and DeepSeek

Six hosted API models and the Evon v3.3 checkpoint scored on identical MILU items with identical prompts. On Indian-language comprehension Evon v3.3 reaches parity with gpt-5.4-nano within 0.23 points and leads DeepSeek-V4-Flash by 6.98 — while being the only model in the comparison alongside DeepSeek that can be deployed inside the customer's own infrastructure.

The efficiency frontier

Accuracy on its own picks the biggest model; cost on its own picks the weakest. Plotting them against each other shows the trade-off a buyer actually faces — and which models are simply dominated, beaten on both axes at once. Switch the horizontal axis to read the same accuracy against price, wait or speed.

Read up and to the left as better on the cost and latency views, up and to the right on output speed. Evon v3.3 is the only model in this comparison that is not dominated on the cost view — no model here is both more accurate and cheaper per point of Indian-language accuracy. gemini-3.5-flash-lite is the accuracy ceiling at 90.41 and gpt-5.4-nano the closest accuracy match at 79.69; both cost materially more per point and neither can be deployed in-VPC.

Where Evon v3.3 wins outright

Three of the six hosted models lead Evon v3.3 on the raw Indic macro — gpt-5.6-luna, gpt-5.4-mini and gemini-3.5-flash-lite. None of them can be self-hosted, and each carries a cost index between 4.7× and 17.5× Evon v3.3's.

Subject-Level Detail

Evon v3.3 traces a hosted frontier model, domain for domain

Grouping MILU's 41 subjects into eight domains shows the shape of the parity result more clearly than the aggregate does. Plotted as a profile, the two models are almost the same shape: Evon v3.3 tracks gpt-5.4-nano within roughly a point on seven of eight domains, and leads outright on Health & Life Sciences and Law & Governance.

Reasoning

Strong at mathematics, in the languages that need it most

Safety & Security

Mid-pack on the overall sweep, first on misinformation

IndicSafe scores refusal and safe-completion behaviour on 5,965 adversarial prompts across 12 Indian languages. Evon v3.3 lands 86.50 overall — 3.27 points behind the strongest model in the sweep. The category breakdown is where the useful signal is: it leads every model in the comparison on misinformation and conspiracy, the harm class that matters most for BFSI and public-sector deployments.

Evon v3.3 leads the field outright on misinformation and conspiracy by 9.50 points, and sits within two points of the best model on four more categories — harmless control, tricky ambiguous, offensive and hate speech, and gender and caste discrimination, where the whole field is within 0.6 of a point. The widest gaps are political manipulation and religious provocation; both are targets for the next safety-tuning pass. Gender and caste discrimination is the hardest category for every model in the sweep — none clears 71.

Interactive Benchmark Explorer

Explore every Indian-language benchmark

Pick a benchmark and toggle comparators to redraw the chart. The complete matrix is always in the table below, so nothing is hidden behind an interaction.

Complete Data

The complete benchmark table

Every benchmark on this page against every comparator with a published or measured score. Filter by category, search by name, click any column to sort.

Values are drawn from Gnani's evaluation harness — Evon v3.3 and any competitor figures Gnani measured directly — or from the first-party and independent published sources listed in References. Gaps of a point or two between models should be read as noise rather than as a stable ranking.

The Mechanism

One Indian word now costs 38.6% fewer tokens

Actual tokenizer output, model by model

Every split below is real tokenizer output, not an illustration — run the sentence through each vendor's own published tokenizer (Hugging Face tokenizer.json for Evon, Sarvam, Gemma, DeepSeek, Llama and Qwen; OpenAI's tiktoken o200k_base for the GPT-5 family) and record exactly where it cuts. Evon v3.3 is pinned at the top of every language; the rest are sorted by token count so you can see exactly how far ahead or behind Evon lands.

Measured on a 1,200-sentence-per-language held-out slice of the AI4Bharat Samanantar corpus (English baseline from the paired source sentences). Sentences above are drawn from that same corpus. Byte fragments occur when a tokenizer's vocabulary has no entry for a script's multi-byte UTF-8 sequence and falls back to splitting inside it — the fragment shown is the raw byte value, not a character, because no character exists at that split point.

Evon v3.3 posts the lowest 8-language average of any tokenizer measured, and the outright lowest fertility on Telugu, Tamil, Gujarati, Kannada and Malayalam. Sarvam-105B's Gemma-4-derived vocabulary is more efficient on Hindi, Bengali, Marathi, Punjabi, Odia and English. The comparison is against each vendor's own published tokenizer files, not a proxy — with one exception: Google has not published Gemini's tokenizer, so Gemma 4's is used as the closest available proxy and is labelled as such throughout. DeepSeek, Llama and Qwen's byte-level BPE tokenizers were not built for Indic scripts and fall back to raw UTF-8 byte splitting on most Indian words, which is why their fertility is 2–7× higher across the board.

Vocabulary coverage, language by language

Punjabi and Odia are two of Evon v3.3's 11 supported languages; they sit outside this vocabulary-extension pass, and the table above shows it — Evon's Punjabi and Odia fertility lands within 0.02% of its own pre-extension tokenizer. Sarvam-105B's tokenizer is currently more efficient on both: 1.63 vs Evon's 3.28 on Punjabi, and 1.89 vs Evon's 18.34 on Odia. Extending coverage to these two languages is the target for the next vocabulary revision.

Serving Performance

Fast, and it stays fast

Evon v3.3 reaches its first token in 302 milliseconds on a short prompt. Give it a hundred thousand tokens of context and that becomes 928 ms — a hundredfold more input for roughly three times the wait. Once generating, it sustains 257.3 tokens per second and barely varies, holding within 7.5% even at 100k input. Predictable is what an SLA is written against.

Every Evon v3.3 figure on this page is measured on Gnani's own inference endpoint with the model running in its default reasoning mode — Evon v3.3 reasons by default, and this is the throughput a production deployment sees today, not a best-case or non-reasoning number. Every comparator figure is Artificial Analysis's own published benchmark data for that model, not Gnani's measurement of it, captured at whatever reasoning configuration Artificial Analysis reports for that model — shown per row in the Configuration column below. ^[7]

On time to first token at this workload Evon v3.3 is second of seven, behind DeepSeek-V4-Flash at 1.37s and ahead of every OpenAI and Google model here. On output speed it is third of seven, ahead of all three OpenAI models and DeepSeek-V4-Flash.

Economics

Frontier-class Indic accuracy at the lowest cost index in the field

Accuracy alone overstates a model's position and cost alone understates it. Dividing one by the other gives the number a buyer actually optimises: what a point of Indian-language accuracy costs. Rebased so Evon v3.3 = 100, every other model in this comparison is at least 2.2× more expensive per point.

How the cost index is derived

The index answers one question: what does a point of Indian-language accuracy cost? For each model, we take its blended per-million-token price — each vendor's published input and output token rates combined in a fixed ratio approximating a typical Indic conversational workload — and divide by that model's Indic MILU macro score on the nine-language comparison set. That gives an absolute cost-per-accuracy-point figure for every model, which we then rebase so Evon v3.3's own figure equals 100; every other model's index is that same ratio, scaled to the same baseline. No absolute prices appear on this page because list prices move faster than a benchmark refresh cycle — the relative index is what survives that churn, and it is only meaningful as a comparison within this set, not as a standalone number.

Deployment

Where Evon v3.3 is production-ready today

Ordered by strength of evidence. Each case draws only on properties measured on this page.

Available today

Get the weights

Evon v3.3 ships as open weights under Apache 2.0. It fits on a single inference node, so it can run inside your own VPC from day one — no data leaves your network to use it.

Repository metadata, licence and architecture values above were read from the published model repository. Confirm the access-approval workflow and any customer-specific licence terms with Gnani before committing to a deployment timeline.

Transparency

How these benchmarks were run

Two different provenances appear on this page, and they are not interchangeable. Indic results are ours, run end to end on our own harness. Global and frontier-model figures are other people's published numbers, cited rather than reproduced.

Reading the numbers

• Evon v3.3 ran on our fleet, hosted comparators through vendor APIs on a different date — read small gaps as noise, not ranking.
• Two MILU aggregates appear and are not interchangeable: the 11-language macro (78.74) and the nine-language comparison macro (79.46). Each is labelled where used.
• Cost index is relative, rebased so Evon v3.3 = 100. It is blended per-million-token price ÷ Indic MILU, never a price or a quotation.
• Tokenizer fertility against Evon's own base is an internal-improvement metric; the cross-model table is a separate, like-for-like comparison.
• Shading is sparse and directional. Differences of a point or two sit inside what these benchmarks resolve.
• Category averages and macros are computed from the per-benchmark values shown, not transcribed separately.

Third-party figures were captured 11–12 August 2026 and pricing-derived values move frequently. Re-verify before external quotation.

FAQs

Frequently asked questions

Sources

References

Third-party figures on this page are published values retrieved 11–12 August 2026. They are not attributable to those vendors as their own published benchmark results, and pricing-derived values move frequently.

1. OpenAI. “Introducing gpt-5.4-mini and gpt-5.4-nano.” openai.com
2. OpenAI. “GPT-5.6 model index.” openai.com
3. DeepSeek AI. “DeepSeek-V4-Flash model card.” huggingface.co
4. Sarvam AI. “Sarvam-30B & Sarvam-105B.” sarvam.ai/blogs
5. Google DeepMind. “Gemini 3.1 Flash-Lite model card.” deepmind.google
6. Google DeepMind. “Gemma 4 Technical Report.” arXiv, 2026. arxiv.org/abs/2607.02770
7. Artificial Analysis. Per-model latency, throughput and pricing pages, retrieved 11–12 August 2026. artificialanalysis.ai
8. BenchmarkList. Independent evaluation index aggregating Vals.ai and Scale Labs results. benchmarklist.com