Indic language models

Most frontier models handle Hindi passably and the rest of India's languages unevenly. Sarvam AI builds specifically for the Indian language landscape, and their models are available on Gopura through the same endpoint, the same key and the same wallet as every other model.

curl https://api.gopura.net/v1/chat/completions \
  -H "Authorization: Bearer $GOPURA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sarvam/sarvam-105b",
    "messages": [
      {"role": "user", "content": "मुंबई में मानसून के मौसम के बारे में दो वाक्य लिखिए।"}
    ]
  }'

The models

Both support tools and structured outputs (response_format, including JSON schema with strict mode). Cached input is billed at $0.120/M.

The two behave very differently, and the per-token rate is only half the story. sarvam-105b-conversations answers directly: a two-sentence Hindi reply costs about $0.00005. sarvam-105b is a heavy reasoner that thinks before it speaks, and you pay for the thinking — answering "what is 2+2" took it 889 output tokens and $0.00075 in our own testing. Reach for the reasoning model when the problem earns it, not by default.

If you send no max_tokens, we fill in 4,096 for sarvam-105b. That number is deliberate: Sarvam's own default of 2,048 is not always enough for the model to finish thinking, and a truncated answer still bills every token it burned. Raising it is not free either — given 8,192 the same arithmetic question stretched to 5,546 tokens. Set max_tokens yourself when you know the shape of your workload; your value always wins.

Sarvam publishes its prices in rupees. We convert at a rate pinned slightly below spot and hold it steady, so your per-token price does not move with the currency — the exact billed cost still arrives in-band on every response, as everywhere else on Gopura.

When to reach for them

Use sarvam-105b when the input or output is substantially in an Indian language and quality matters: Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia and Assamese, alongside English. Code-mixed text — Hinglish and its regional cousins — is the case where the gap against general-purpose models is widest.

Reach for sarvam-105b-conversations when latency is what you are optimising: support chat, IVR, and voice agents where a reply has to land in a beat. For most Indic-language work it is the better first choice — it is faster, far cheaper per answer, and its language quality is excellent.

The reasoning model's chain of thought arrives as reasoning on the wire, the same field every other reasoning model on Gopura uses.

For everything else, the general catalog is usually the better pick. These are specialists, not replacements, and gopura/auto will not route to them — name them explicitly when you want them.

Pairing with the India document pack

Sarvam's language handling and our India document pack solve adjacent halves of the same problem. Extract a PAN card or a GST certificate with /v1/docs/in/..., then reason about the result in the applicant's own language:

curl https://api.gopura.net/v1/chat/completions \
  -H "Authorization: Bearer $GOPURA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "sarvam/sarvam-105b",
    "messages": [{
      "role": "user",
      "content": "इस GST प्रमाणपत्र के आधार पर बताइए कि पंजीकरण किस राज्य में है: {\"gstin\": \"27AAPFU0939F1ZV\"}"
    }]
  }'

On data residency

We do not claim that requests to these models stay within India. Sarvam publishes no residency guarantee we can pass on, and Gopura's own processing runs at the network edge rather than pinned to a region — so a residency claim would be marketing rather than fact. What we do guarantee is what we guarantee everywhere: we never log prompts or completions.

If enforced regional processing is a requirement for you, tell us — it is on the roadmap, and it needs infrastructure commitments we would rather make against real demand than in a brochure.