Model card

Sarvam 105B

@sdn/sarvam-105bGeneral purposeLaunch planLive in the catalog

Sarvam AI's flagship model, released with open weights under Apache 2.0. It is a mixture-of-experts model with about 10B of its 106B parameters active per token, trained for 22 Indian languages plus English, so it handles Hindi, Hinglish and regional-language prompts well. Reasoning is on by default. Send reasoning_effort none to turn it off, and the first token arrives in about 0.6 s. In our tests it called tools reliably, including parallel calls, and with reasoning off it finished a coding fix that spanned several files.

Context window128K tokens

Served bySideren engine

Where it earns its keep[01/03]

0122 Indian languages and English, Hinglish included
02Reliable tool calls, including parallel calls
03Reasoning you can switch off for first tokens under a second

Capabilities

Tool callingYes

Strict, schema-faithful tool calls, enforced by the engine on every request — safe to build an agent loop on.

ReasoningYes

Thinks before it answers. Budget max_tokens generously — hidden reasoning counts against it.

VisionNo

Text-only.

Behind the endpoint[02/03]

One endpoint. Served by our engine.

Sarvam 105B is served through the Sideren engine — zero-downtime serving is the design target, not a status-page apology. You request it by name; everything else is our problem.

Call it by name

curl https://api.sideren.io/v1/messages \
  -H "x-api-key: sdn_your_key" \
  -H "content-type: application/json" \
  -d '{
    "model": "@sdn/sarvam-105b",
    "max_tokens": 1024,
    "messages": [{ "role": "user", "content": "Hello" }]
  }'

OpenAI-style clients work too — send the same model name to /v1/chat/completions with a Bearer key. See the docs for both dialects.

Your agent stays the same.
Its inference gets better.

$export ANTHROPIC_BASE_URL=https://api.sideren.io

Get a key Read the docs

Start free · no card · Paid plans from $5/mo

One endpoint. Served by our engine.

Your agent stays the same.Its inference gets better.

Your agent stays the same.
Its inference gets better.