Skip to navigation

Sarvam-105B

View as Markdown

Sarvam-105B (Flagship Chat LLM)

Sarvam AI’s flagship Mixture-of-Experts reasoning model trained from scratch, with Multi-head Latent Attention (MLA) for efficient long-context inference. Matches or outperforms most open and closed-source frontier models of its class across knowledge, reasoning, and agentic benchmarks.

Highlights:

  • 105B+ total parameters: our most capable MoE model with Multi-head Latent Attention
  • Pre-trained on 12 trillion tokens across code, math, multilingual, and web data
  • 98.6 on Math500, 88.3 on AIME 25 (96.7 with tools), 49.5 on BrowseComp
  • State-of-the-art Indian language performance: wins 90% of pairwise comparisons
  • Powers Indus, Sarvam’s AI assistant for complex reasoning and agentic workflows
  • OpenAI-compatible chat completions API | Apache 2.0 open-source

At a Glance

Model IDsarvam-105b (flagship) · sarvam-105b-conversations (conversational)
What it doesFlagship chat LLM, complex reasoning, coding, and agentic workflows (105B+ MoE with MLA). The conversational variant is post-trained for real-time dialogue and voice-agent use cases.
Languages10 most-spoken Indian languages + English; native script, romanized, and code-mixed input
APIsChat Completions (POST /v1/chat/completions; sarvam-105b also on /v2/chat/completions, OpenAI-compatible, streaming supported)
Input limits128K-token context window (sarvam-105b); 32K-token context window (sarvam-105b-conversations), all limits
BenchmarksKey Features and the model blog
Pricing₹10 input · ₹0.25 cached input · ₹25 output per 1M tokens, introductory pricing through the end of 2026 (Pricing page)
Best forMulti-step reasoning, code generation, long-context document analysis, agentic tool use
Known limitationsSee below

Key Features

Flagship Indian Language Support

Wins 90% of pairwise comparisons across Indian language benchmarks and 84% on STEM, math, and coding. Trained extensively on native script, romanized, and code-mixed inputs across the 10 most-spoken Indian languages.

Advanced Reasoning

98.6 on Math500, 88.3 on AIME 25 (96.7 with tools), 85.8 on HMMT, and 69.1 on Beyond AIME, reflecting deep multi-step reasoning and complex mathematical problem solving.

Agentic Capabilities

49.5 on BrowseComp and 68.3 on Tau2 (avg.), highest among compared models. Optimized for tool use, long-horizon reasoning, and environment interaction in real-world workflows.

Efficient MoE Architecture

Mixture-of-Experts Transformer with 128 sparse experts and Multi-head Latent Attention (MLA), a compressed attention formulation that reduces memory requirements for long-context inference.

Learn More

For detailed information on architecture, training methodology, performance benchmarks, and inference optimizations, visit our blog.

Model Specifications

Key Considerations
  • Model IDs: sarvam-105b, sarvam-105b-conversations
  • Total Parameters: 105B+ with MoE architecture and 128 sparse experts
  • Attention: Multi-head Latent Attention (MLA)
  • Pre-training Data: 12T tokens
  • Temperature range: 0 to 2
  • Top-p range: 0 to 1
  • Supports streaming and non-streaming responses
  • OpenAI-compatible chat completions format
  • License: Apache 2.0

Sarvam-105B is our flagship chat model, delivering the highest quality outputs for complex reasoning and agentic tasks. Use model ID sarvam-105b-conversations when you need a conversational variant tuned for real-time dialogue and voice agents. See Conversational variant. Sarvam-M (24B) has been deprecated and is no longer available through the API.

Conversational variant

Model ID sarvam-105b-conversations is a post-trained variant of Sarvam-105B for real-time conversational workloads: voice agents, chatbots, and multi-turn dialogue where natural, colloquial Indic responses matter more than deep reasoning traces.

sarvam-105bsarvam-105b-conversations
Best forComplex reasoning, coding, long-context analysis, agentic tool useReal-time dialogue, voice agents, customer-facing chat
EndpointPOST /v1/chat/completions or POST /v2/chat/completionsPOST /v1/chat/completions only
Context window128K tokens32K tokens
ReasoningOn by default; turn it off with reasoning_effort=NoneAlways off; a reasoning_effort level is accepted but ignored
Default temperature1.0 with reasoning on, 0.6 with reasoning off0.6
PricingSarvam 105BSarvam 105B Chat, same rates. Introductory pricing through the end of 2026

Both model IDs share the same OpenAI-compatible request schema and Indian language coverage, but sarvam-105b-conversations has a smaller 32K context window (vs. 128K for sarvam-105b), and it runs with reasoning off only. It doesn’t support thinking mode: if you send a reasoning_effort level, the request succeeds but the model doesn’t reason, so you can leave the field out. Pass the model ID in the model field; the rest of the request stays the same.

from sarvamai import SarvamAI
client = SarvamAI(api_subscription_key="YOUR_SARVAM_API_KEY")
response = client.chat.completions(
model="sarvam-105b-conversations",
messages=[
{"role": "user", "content": "who are you?"}
],
temperature=0.6,
top_p=1,
max_tokens=100,
)
print(response.choices[0].message.content)

Key Capabilities

Simple, one-turn interaction where the user asks a question and the model replies with the highest quality response leveraging its 105B parameter knowledge.

from sarvamai import SarvamAI
client = SarvamAI(
api_subscription_key="YOUR_SARVAM_API_KEY",
)
response = client.chat.completions(
model="sarvam-105b",
messages=[
{"role": "user", "content": "Explain the economic impact of GST implementation in India in about 200 words."}
],
temperature=0.5,
top_p=1,
max_tokens=6000,
)
print(response.choices[0].message.content)

Limits

LimitValue
Context window128K tokens (sarvam-105b) / 32K tokens (sarvam-105b-conversations)
max_tokensStarter 4096 / Pro 16384 / Business 128000. When omitted on /v1/chat/completions, it defaults to 2,048, and reasoning counts toward it
temperature0–2. sarvam-105b: default 1.0 with reasoning on (the default), 0.6 with reasoning off. sarvam-105b-conversations: default 0.6 (reasoning is always off)
top_p0–1
n (completions per request)1 only; higher values return 400
frequency_penalty / presence_penalty-2 to 2
stopUp to 4 sequences
Rate limitsSee Rate Limits

Known Limitations

LimitationDetailWorkaround
Thinking mode is on by defaultOn sarvam-105b, reasoning (reasoning_effort, default low) is enabled by default, and reasoning tokens count toward completion tokens, a small max_tokens can be consumed entirely by reasoningIncrease max_tokens to leave room for the visible answer, or disable reasoning with reasoning_effort=None

Next Steps