Gemma 4 31B | Sarvam API Docs

Gemma 4 31B is an open-source model served by Sarvam in beta, and is not tuned for Indian languages. Beta access is granted per API key — contact us to request whitelisting. For Indian language workloads use Sarvam-105B; for document and image understanding in Indian languages, use Sarvam Vision.

A 31B-parameter instruction-tuned model. Its distinguishing feature on Sarvam is image input — it is the only chat model here that accepts images alongside text in the chat completions format.

At a Glance


Model ID	`gemma4`
What it does	Chat LLM with image understanding
Languages	Not tuned for Indian languages by Sarvam
APIs	Open-Source Models API Reference (`POST /v2/chat/completions`, OpenAI-compatible, streaming supported) — also see using an open-source model
Input limits	131,072-token context window; images as base64 data URIs only; 10 MB request cap — all limits
Benchmarks	None published by Sarvam
Pricing	₹36.6 input · ₹13.73 cached input · ₹91.5 output — per 1M tokens (Pricing)
Best for	Image captioning, classification, and visual question answering
Not for	Indian language workloads or document OCR — use Sarvam-105B or Sarvam Vision

Why you might use it

Image understanding

Accepts images inline as base64 data URIs in the standard image_url content part, so you can caption, classify, or answer questions about images.

Direct answers

Not a reasoning model — it replies without a chain-of-thought pass, so the whole token budget goes to the answer.

Model Specifications

Key Considerations

Model ID: gemma4
Context window: 131,072 tokens (128K)
Image input: supported — base64 data URIs only
Tool calling: supported
Reasoning: none — answers directly, never returns reasoning_content, and reasoning_effort has no effect
Maximum request size: 10 MB (inclusive of encoded images)
Temperature range: 0 to 2 (unset uses the model’s own default)
Top-p range: greater than 0, up to 1
Supports streaming and non-streaming responses
OpenAI-compatible chat completions format

Each image is estimated at roughly 1,024 tokens when Sarvam checks your request against the context window. Budget for this when sending several images with a long prompt.

Gemma 4 31B vs other options

Feature	Gemma 4 31B	GLM-5.2	Sarvam-105B	Sarvam Vision
Context window	131,072	524,288	128,000	—
Image input	✅	❌	❌	✅ (documents)
Tool calling	✅	✅	✅	—
Indian language tuning	None	None	23 languages	23 languages
Best for	Image + text tasks	Very long-context tasks	Indian-language chat & agentic work	Document OCR & extraction

For document understanding — scanned forms, invoices, PDFs, especially in Indian languages — use Sarvam Vision, which is purpose-built for OCR and structured extraction. Use gemma4 for general images where you want a conversational answer rather than structured document output.

Key Capabilities

Basic Chat Completion

Image Input

Streaming

A text-only single-turn request. Only the model field differs from a Sarvam chat completion call.

1 from sarvamai import SarvamAI
2 
3 client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY")
4 
5 response = client.open_source_models.chat_completions_v2(
6     model="gemma4",
7     messages=[
8         {"role": "user", "content": "Classify this support ticket as billing, technical, or other: 'My invoice shows twice the usage I expected.'"}
9     ],
10     temperature=0.1,
11     max_tokens=100,
12 )
13 
14 print(response.choices[0].message.content)

Parameters

Gemma 4 31B accepts the standard OpenAI-compatible parameter set — temperature, top_p, max_tokens, stream, stop, n, seed, frequency_penalty, presence_penalty, response_format (Structured Outputs), tools, and tool_choice. See the full parameter table for types, defaults, and ranges.

Model-specific notes:

Images must be base64 data URIs — remote http(s) URLs are rejected.
No reasoning — the model answers directly, never returns reasoning_content, and reasoning_effort is accepted but has no effect.
seed is reproducible — identical seed and parameters return identical output in current serving.
Don’t ask for JSON in the prompt when response_format is set — an instruction like “Return JSON” alongside a JSON response format makes the model emit whitespace until it exhausts max_tokens, billing the full budget for no usable output. Let response_format handle the shape.

Limits and errors

Common errors for every model on /v2/chat/completions are documented on the Open-Source Models overview. Gemma-4-specific cases:

Condition	Status	Code	What it means
API key lacks beta access	`400`	`invalid_request_error`	Request whitelisting — see Beta APIs.
Missing or invalid API key	`403`	`invalid_api_key_error`	Check the `api-subscription-key` header.
No credits remaining	`402`	`insufficient_quota_error`	Top up on the dashboard.
Unknown or unavailable model id	`404`	`not_found_error`	Confirm the id with `GET /v2/models`.
`image_url` pointing at an `http(s)` URL	`400`	`invalid_request_error`	Inline the image as a base64 data URI instead.
`max_tokens` above 131,072	`400`	`invalid_request_error`	`max_tokens` cannot exceed the context window.
Estimated prompt tokens + `max_tokens` above 131,072	`422`	`unprocessable_entity_error`	Shorten the prompt, send fewer images, or lower `max_tokens`.
Request body above the model’s size cap	`413`	`invalid_request_error`	Encoded images count toward this cap.
Invalid tool schema	`400`	`invalid_request_error`	See Tool schemas.
Whitespace-only reply with `finish_reason: "length"`	`200`	—	`response_format` was combined with a “return JSON” instruction in the prompt — remove the instruction.

Support

Sarvam supports the serving layer — auth, billing, rate limits, availability. Sarvam does not tune or evaluate this model, so its output quality, reasoning, and language coverage are properties of the model itself. See Open-Source Models for the full support boundary, and Talk to us for serving issues.