105B
- Context
- 33Ktokens
- Released
- 1y ago
- Size
- 23.6Bparameters
Sarvam-M is a multilingual language model from Sarvam AI, post-trained from Mistral Small (M stands for Mistral) using supervised fine-tuning and reinforcement learning with verifiable rewards. Sarvam AI reports an average 20% improvement on Indian language benchmarks over the base model, with further gains in math and programming. It is intended for reasoning-heavy uses such as math, coding, and everyday conversation in Indian languages and English, offering both a thinking and a non-thinking mode.
How it compares
Additional source evidence
- Source model
- sarvam/sarvam-m
- Source family
- sarvam
- Last updated
- 2025-05-20
- Model context
- 33K
- Model output limit
- 33K
- Open weights
- Yes
- File attachments
- No
- Reasoning
- Yes
- Tool calling
- No
- Temperature control
- Yes
Downloadable weights
Complete source record (JSON)
{
"id": "sarvam/sarvam-m",
"name": "Sarvam-M",
"description": "Hybrid-reasoning multilingual model for Indian languages built on Mistral Small",
"family": "sarvam",
"last_updated": "2025-05-20",
"attachment": false,
"reasoning": true,
"tool_call": false,
"temperature": true,
"modalities": {
"input": [
"text"
],
"output": [
"text"
]
},
"limit": {
"context": 32768,
"output": 32768
},
"release_date": "2025-05-20",
"open_weights": true,
"license": "Apache 2.0",
"weights": [
{
"label": "Hugging Face",
"url": "https://huggingface.co/sarvamai/sarvam-m"
}
]
}Frequently asked questions
Running locally
Sarvam-M is estimated to need 16.9 GB of VRAM at Q4, 29.3 GB at Q8 and 54.7 GB at FP16. These are estimates based on the 23.6B parameter count, with weights plus about 15% and 0.5 GB for a short context.
Yes, it fits on a 24 GB GPU at Q4 (estimated 16.9 GB). It does not fit on a 16 GB, 12 GB or 8 GB GPU at Q4, Q8 or FP16 (estimate).
Yes. Assuming macOS lets the GPU use about 75% of unified memory, the smallest common Mac memory size that fits the estimated 16.9 GB at Q4 is 24 GB.
Privacy and data
Yes, Sarvam-M has open weights, so you can self-host it for full control of your data.
Capabilities and specs
The context window is 33K tokens (32,768) and the maximum output is 33K tokens (32,768).
No, Sarvam-M only accepts text as input. It cannot read images, audio, video or PDFs.
No, Sarvam-M does not support tool calling (function calling).
Yes, Sarvam-M supports a reasoning mode.
Quality and benchmarks
The closest alternatives, matched on model kind and release date with no comparable price, are FastVLM 0.5B by Apple, FastVLM 1.5B by Apple and Qwen3 Embedding 8B by Alibaba.
FastVLM 0.5BSarvam-M vs FastVLM 0.5BFastVLM 1.5BSarvam-M vs FastVLM 1.5BQwen3 Embedding 8BSarvam-M vs Qwen3 Embedding 8B
Licensing and access
Sarvam-M is open weights, released under the Apache 2.0 license.
Yes, Sarvam-M is licensed under Apache 2.0, a permissive license that allows commercial use.
The Sarvam-M weights can be downloaded from the Hugging Face repository sarvamai/sarvam-m.
Basics
Sarvam-M was made by Sarvam AI and released on May 20, 2025.
Yes, Sarvam-M has open weights.
Generated from Vertical Tabs data, updated November 4, 2026.
Discussion
No comments yet
Be the first to share your thoughts.