567-labs · EshwarCVS · Jul 9, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -7,7 +7,14 @@ Versioning: [Semantic Versioning](https://semver.org/spec/v2.0.0.html)
 
 ---
 
-## [1.15.5] - 2026-06-28
+## [Unreleased]
+
+### Added
+- **Sarvam**: First-class `from_sarvam()` and `from_provider("sarvam/...")` support for Sarvam's OpenAI-compatible Indic chat API (`sarvam-30b`, `sarvam-105b`).
+
+### Changed
+- **Sarvam**: Default structured-output mode is `Mode.MD_JSON`.
+---
 
 ### Fixed
 - **v2 imports**: Defer OpenAI SDK imports from core v2 modules until an OpenAI-specific path actually needs them, reducing import side effects for non-OpenAI usage. ([#2390](https://github.com/567-labs/instructor/pull/2390))

diff --git a/docs/concepts/from_provider.md b/docs/concepts/from_provider.md
@@ -73,6 +73,7 @@ user = client.create(
 - Cohere: `"cohere/command-r-plus"`
 - Perplexity: `"perplexity/llama-3.1-sonar"`
 - DeepSeek: `"deepseek/deepseek-chat"`
+- Sarvam: `"sarvam/sarvam-30b"` (Indic languages — Hindi, Tamil, Bengali, and more)
 - xAI: `"xai/grok-beta"`
 - OpenRouter: `"openrouter/meta-llama/llama-3.1-70b"`
 - Ollama: `"ollama/llama3.2"` (local models)

diff --git a/docs/integrations/index.md b/docs/integrations/index.md
@@ -30,6 +30,7 @@ Learn how to integrate Instructor with various AI model providers. These compreh
     [:octicons-arrow-right-16: Cohere](./cohere.md)          ·
     [:octicons-arrow-right-16: Mistral](./mistral.md)        ·
     [:octicons-arrow-right-16: DeepSeek](./deepseek.md)      ·
+    [:octicons-arrow-right-16: Sarvam](./sarvam.md)        ·
     [:octicons-arrow-right-16: Together AI](./together.md)    ·
     [:octicons-arrow-right-16: Groq](./groq.md)              ·
     [:octicons-arrow-right-16: Fireworks](./fireworks.md)    ·

diff --git a/docs/integrations/sarvam.md b/docs/integrations/sarvam.md
@@ -0,0 +1,118 @@
+---
+title: "Structured outputs with Sarvam AI"
+description: "Use Instructor with Sarvam Indic LLMs for type-safe structured extraction across Indian languages."
+---
+
+# Structured outputs with Sarvam AI
+
+[Sarvam AI](https://www.sarvam.ai/models) provides Indic-focused LLMs such as **Sarvam 30B** and **Sarvam 105B** through an [OpenAI-compatible chat API](https://docs.sarvam.ai/). Instructor now supports Sarvam as a first-class provider via `from_provider("sarvam/...")` and `from_sarvam()`.
+
+## Install
+
+```bash
+pip install "instructor[openai]"
+```
+
+Set your API key:
+
+```bash
+export SARVAM_API_KEY='your-api-key-here'
+```
+
+## Quick start
+
+```python
+import instructor
+from pydantic import BaseModel
+
+
+class User(BaseModel):
+    name: str
+    age: int
+
+
+client = instructor.from_provider("sarvam/sarvam-30b")
+
+user = client.chat.completions.create(
+    model="sarvam-30b",
+    response_model=User,
+    messages=[{"role": "user", "content": "राहुल की उम्र 28 साल है।"}],
+)
+
+print(user)
+# > User(name='Rahul', age=28)
+```
+
+## Supported models
+
+Chat LLMs exposed via the Sarvam API (see [Sarvam models](https://www.sarvam.ai/models)):
+
+| Model ID | Description |
+|----------|-------------|
+| `sarvam-30b` | Efficient MoE chat model for high-throughput Indic workloads |
+| `sarvam-105b` | Flagship chat model for highest-quality Indic understanding |
+
+Per [Sarvam's model guide](https://docs.sarvam.ai/api-reference-docs/getting-started/models), chat models support **11 languages** (10 Indian + English). `sarvam-m` is deprecated — use `sarvam-30b` or `sarvam-105b`.
+
+Speech (Saaras), TTS (Bulbul), translation (Mayura, Sarvam Translate), and document intelligence (Sarvam Vision) use separate APIs and are not covered by this integration yet. See [Extending Sarvam support](#extending-sarvam-support).
+
+## Modes
+
+Sarvam works with Instructor's OpenAI-compatible modes:
+
+| Mode | When to use |
+|------|-------------|
+| `Mode.MD_JSON` (default) | Most reliable across Indic scripts in our Jul 2026 benchmark |
+| `Mode.TOOLS` | Fastest latency; best on `sarvam-105b`, but unreliable on `sarvam-30b` for extraction |
+| `Mode.JSON_SCHEMA` | Native `response_format` JSON schema |
+
+!!! warning "`Mode.TOOLS` on `sarvam-30b`"
+    In the full 432-trial eval, `sarvam-30b + Mode.TOOLS` scored **29–46%** success on person extraction (vs **58–75%** for `Mode.MD_JSON`). Prefer `Mode.MD_JSON` for cost-sensitive extraction workloads.
+
+```python
+import instructor
+
+client = instructor.from_provider(
+    "sarvam/sarvam-105b",
+    mode=instructor.Mode.JSON_SCHEMA,
+)
+```
+
+Sarvam-specific API parameters such as `reasoning_effort` and `wiki_grounding` can be passed through `create()` kwargs.
+
+## Async usage
+
+```python
+import asyncio
+import instructor
+from pydantic import BaseModel
+
+
+class CityInfo(BaseModel):
+    city: str
+    state: str
+
+
+async def main() -> None:
+    client = instructor.from_provider("sarvam/sarvam-30b", async_client=True)
+    result = await client.chat.completions.create(
+        model="sarvam-30b",
+        response_model=CityInfo,
+        messages=[{"role": "user", "content": "Extract city info: Bengaluru is in Karnataka."}],
+    )
+    print(result)
+
+
+asyncio.run(main())
+```
+
+## Extending Sarvam support
+
+Future non-chat Sarvam products (Vision, Saaras STT, Bulbul TTS, Translate) would need dedicated provider handlers under `instructor/v2/providers/sarvam/` with separate API endpoints. The current integration is intentionally scoped to chat completions.
+
+## Related resources
+
+- [Sarvam models](https://www.sarvam.ai/models)
+- [Sarvam API documentation](https://docs.sarvam.ai/)
+- [Instructor modes comparison](../concepts/modes-comparison.md)
+- [from_provider guide](../concepts/from_provider.md)
diff --git a/instructor/__init__.py b/instructor/__init__.py
@@ -53,6 +53,7 @@
         from_anyscale as from_anyscale,
         from_databricks as from_databricks,
         from_deepseek as from_deepseek,
+        from_sarvam as from_sarvam,
         from_openai as from_openai,
         from_together as from_together,
     )
@@ -77,6 +78,7 @@
     "from_together",
     "from_databricks",
     "from_deepseek",
+    "from_sarvam",
     "from_openrouter",
     "from_litellm",
     "from_vertexai",
@@ -116,6 +118,7 @@
     "from_together": (".v2.providers.openai.client", "from_together"),
     "from_databricks": (".v2.providers.openai.client", "from_databricks"),
     "from_deepseek": (".v2.providers.openai.client", "from_deepseek"),
+    "from_sarvam": (".v2.providers.openai.client", "from_sarvam"),
     "from_openrouter": (".v2.providers.openrouter.client", "from_openrouter"),
     "from_litellm": (".v2.providers.litellm.client", "from_litellm"),
     "Mode": (".mode", "Mode"),

diff --git a/instructor/models.py b/instructor/models.py
@@ -159,5 +159,8 @@
         # DeepSeek Models
         "deepseek/deepseek-chat",
         "deepseek/deepseek-reasoner",
+        # Sarvam Models - https://docs.sarvam.ai/api-reference-docs/getting-started/models
+        "sarvam/sarvam-30b",
+        "sarvam/sarvam-105b",
     ],
 )
diff --git a/instructor/v2/README.md b/instructor/v2/README.md
@@ -1142,7 +1142,7 @@ This checklist tracks which providers have been migrated to v2:
   - Tests: `tests/v2/test_provider_modes.py`, `tests/v2/test_handlers_parametrized.py`
   - Status: ✅ Complete
 
-- [x] **OpenAI-Compatible** (`Provider.ANYSCALE`, `Provider.TOGETHER`, `Provider.DATABRICKS`, `Provider.DEEPSEEK`)
+- [x] **OpenAI-Compatible** (`Provider.ANYSCALE`, `Provider.TOGETHER`, `Provider.DATABRICKS`, `Provider.DEEPSEEK`, `Provider.SARVAM`)
   - Location: `instructor/v2/providers/openai/`
   - Modes: `TOOLS`, `JSON`, `JSON_SCHEMA`, `MD_JSON`, `PARALLEL_TOOLS`
   - Tests: `tests/v2/test_handlers_parametrized.py`, `tests/v2/test_client_unified.py`

diff --git a/instructor/v2/auto_client.py b/instructor/v2/auto_client.py
@@ -1413,6 +1413,68 @@ def _build_deepseek(
         raise
 
 
+def _build_sarvam(
+    *,
+    provider: str,
+    model_name: str,
+    async_client: bool,
+    mode: Mode | None,
+    api_key: str | None,
+    kwargs: dict[str, Any],
+    provider_info: dict[str, str],
+) -> InstructorType:
+    try:
+        import openai
+        from instructor.v2.providers.openai.client import from_sarvam
+        import os
+
+        api_key = api_key or os.environ.get("SARVAM_API_KEY")
+
+        if not api_key:
+            from instructor.v2.core.errors import ConfigurationError
+
+            raise ConfigurationError(
+                "SARVAM_API_KEY is not set. "
+                "Set it with `export SARVAM_API_KEY=<your-api-key>` or pass it as kwarg api_key=<your-api-key>"
+            )
+
+        base_url = kwargs.pop("base_url", "https://api.sarvam.ai/v1")
+
+        client = (
+            openai.AsyncOpenAI(api_key=api_key, base_url=base_url)
+            if async_client
+            else openai.OpenAI(api_key=api_key, base_url=base_url)
+        )
+
+        result = from_sarvam(
+            client,
+            model=model_name,
+            mode=mode if mode else Mode.MD_JSON,
+            **kwargs,
+        )
+        logger.info(
+            "Client initialized",
+            extra={**provider_info, "status": "success"},
+        )
+        return result
+    except ImportError:
+        from instructor.v2.core.errors import ConfigurationError
+
+        raise ConfigurationError(
+            "The openai package is required to use the Sarvam provider. "
+            "Install it with `pip install openai`."
+        ) from None
+    except Exception as e:
+        logger.error(
+            "Error initializing %s client: %s",
+            provider,
+            e,
+            exc_info=True,
+            extra={**provider_info, "status": "error"},
+        )
+        raise
+
+
 def _build_xai(
     *,
     provider: str,
@@ -1602,6 +1664,7 @@ def _build_litellm(
     "generative-ai": _build_generative_ai,
     "ollama": _build_ollama,
     "deepseek": _build_deepseek,
+    "sarvam": _build_sarvam,
     "xai": _build_xai,
     "openrouter": _build_openrouter,
     "litellm": _build_litellm,

diff --git a/instructor/v2/core/mode.py b/instructor/v2/core/mode.py
@@ -75,6 +75,10 @@ class Mode(enum.Enum):
     PERPLEXITY_JSON = "perplexity_json"
     OPENROUTER_STRUCTURED_OUTPUTS = "openrouter_structured_outputs"
 
+    # Sarvam modes
+    SARVAM_TOOLS = "sarvam_tools"
+    SARVAM_JSON = "sarvam_json"
+
     # Classification helpers
     @classmethod
     def tool_modes(cls) -> set["Mode"]:
@@ -102,6 +106,7 @@ def tool_modes(cls) -> set["Mode"]:
             cls.GENAI_TOOLS,
             cls.RESPONSES_TOOLS,
             cls.RESPONSES_TOOLS_WITH_INBUILT_TOOLS,
+            cls.SARVAM_TOOLS,
         }
 
     @classmethod
@@ -124,6 +129,7 @@ def json_modes(cls) -> set["Mode"]:
             cls.OPENROUTER_STRUCTURED_OUTPUTS,
             cls.MISTRAL_STRUCTURED_OUTPUTS,
             cls.XAI_JSON,
+            cls.SARVAM_JSON,
         }
 
     @classmethod
@@ -247,6 +253,9 @@ def warn_deprecated_mode(cls, mode: "Mode") -> None:
     Mode.GEMINI_JSON: Mode.MD_JSON,
     # OpenRouter legacy modes
     Mode.OPENROUTER_STRUCTURED_OUTPUTS: Mode.JSON_SCHEMA,
+    # Sarvam legacy modes
+    Mode.SARVAM_TOOLS: Mode.TOOLS,
+    Mode.SARVAM_JSON: Mode.MD_JSON,
 }
 
 

diff --git a/instructor/v2/core/provider_specs.py b/instructor/v2/core/provider_specs.py
@@ -145,6 +145,23 @@ def _openai_compat_spec(
             alias="deepseek",
             from_function="from_deepseek",
         ),
+        Provider.SARVAM: _spec(
+            Provider.SARVAM,
+            aliases=("sarvam",),
+            handler_module="instructor.v2.providers.openai.handlers",
+            supported_modes=(Mode.TOOLS, Mode.JSON, Mode.JSON_SCHEMA, Mode.MD_JSON),
+            unsupported_modes=(Mode.PARALLEL_TOOLS, Mode.RESPONSES_TOOLS),
+            legacy_modes={
+                Mode.SARVAM_TOOLS: Mode.TOOLS,
+                Mode.SARVAM_JSON: Mode.MD_JSON,
+            },
+            from_function="from_sarvam",
+            client_module="instructor.v2.providers.openai.client",
+            sdk_module="openai",
+            provider_string="sarvam/sarvam-30b",
+            basic_modes=(Mode.TOOLS, Mode.JSON_SCHEMA, Mode.MD_JSON),
+            async_modes=(Mode.TOOLS, Mode.JSON_SCHEMA, Mode.MD_JSON),
+        ),
         Provider.OPENROUTER: _spec(
             Provider.OPENROUTER,
             aliases=("openrouter",),

diff --git a/instructor/v2/core/providers.py b/instructor/v2/core/providers.py
@@ -36,6 +36,7 @@ class Provider(Enum):
     BEDROCK = "bedrock"
     PERPLEXITY = "perplexity"
     OPENROUTER = "openrouter"
+    SARVAM = "sarvam"
 
 
 def provider_from_mode(mode: Mode, default: Provider = Provider.OPENAI) -> Provider:
@@ -69,6 +70,8 @@ def provider_from_mode(mode: Mode, default: Provider = Provider.OPENAI) -> Provi
         Mode.BEDROCK_JSON: Provider.BEDROCK,
         Mode.PERPLEXITY_JSON: Provider.PERPLEXITY,
         Mode.OPENROUTER_STRUCTURED_OUTPUTS: Provider.OPENROUTER,
+        Mode.SARVAM_TOOLS: Provider.SARVAM,
+        Mode.SARVAM_JSON: Provider.SARVAM,
     }
     return mapping.get(mode, default)
 
@@ -119,6 +122,7 @@ def get_provider(base_url: str) -> Provider:
         ("openrouter", Provider.OPENROUTER),
         ("x.ai", Provider.XAI),
         ("xai", Provider.XAI),
+        ("sarvam", Provider.SARVAM),
     )
     for token, provider in providers:
         if token in normalized: