README.md

---
library_name: transformers
license: apache-2.0
language:
- en
- bn
- hi
- kn
- gu
- mr
- ml
- or
- pa
- ta
- te
base_model:
- mistralai/Mistral-Small-3.1-24B-Instruct-2503
---

## Model Information

Sarvam-M multilingual hybrid reasoning llm is an instruction tuned generative model in 24B (text in/text out) post trained over Mistral 3.1 24B. It significantly improves on the base Mistral model: +20% average improvement on Indian language benchmarks, +21.6% on math benchmarks, and +17.6% on programming benchmarks. The gains in tasks in the intersectionality of Indian languages and math are even higher, e.g., +86% improvement in a romanized Indian language GSM-8K benchmark.

Learn in detail about sarvam-M in our [blog post](link)

## Key Features
- **Hybrid thinking mode** A single model supports both "think" and "non-think" modes. Use the think mode for tasks requiring complex logical reasoning, math, and coding, and switch to the non-think mode for efficient, general-purpose conversation.
- **Indic Skills** Specifically post-trained on Indian languages alongside English, the model also embodies a character that reflects and emphasizes Indian cultural values.
- **Reasoning capabilities** Sarvam-M outperforms most models of similar size on coding and math benchmarks, demonstrating strong reasoning capabilities.
- **Chatting Experience** With support for both Indic scripts and romanized versions of Indian languages, Sarvam-M offers a smooth and accessible multilingual chat experience.

## Quickstart 

The following contains a code snippet illustrating how to use the model generate content based on given inputs.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "sarvamai/sarvam-M"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name, torch_dtype="auto", device_map="auto"
)

# prepare the model input
prompt = "Who are you and what is your purpose on this planet?"

messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    enable_thinking=True,  # Switches between thinking and non-thinking modes. Default is True.
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(**model_inputs, max_new_tokens=8192)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]) :].tolist()
output_text = tokenizer.decode(output_ids)

if "</think>" in output_text:
    reasoning_content = output_text.split("</think>")[0].rstrip("\n")
    content = output_text.split("</think>")[-1].lstrip("\n").rstrip("</s>")
else:
    reasoning_content = ""
    content = output_text.rstrip("</s>")

print("reasoning content:", reasoning_content)
print("content:", content)
```

## VLLM Deployment

For deployment, you can use `vllm>=0.8.5` to create an OpenAI-compatible API endpoint:
```shell
vllm serve sarvamai/sarvam-M
```

For inference and switching between thinking and non-thinking mode, refer to the below python code:
```python
from openai import OpenAI

# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

models = client.models.list()
model = models.data[0].id

messages = [{"role": "user", "content": "How many letter r in word strawberry?"}]

# By default, the model is in thinking mode.
# If you want to disable thinking, add:
# extra_body={"chat_template_kwargs": {"enable_thinking": False}}
response = client.chat.completions.create(model=model, messages=messages)
output_text = response.choices[0].message.content

if "</think>" in output_text:
    reasoning_content = output_text.split("</think>")[0].rstrip("\n")
    content = output_text.split("</think>")[-1].lstrip("\n").rstrip("</s>")
else:
    reasoning_content = ""
    content = output_text.rstrip("</s>")

print("reasoning content:", reasoning_content)
print("content:", content)

# For the next round, add the assistant's response and reasoning to the messages.
messages.append(
    {"role": "assistant", "content": content, "reasoning_content": reasoning_content}
)
```

The above example also shows how to add assistant turns in the messages for multiturn conversation.