AI Models & Platforms

Sarvam AI Releases Open-Weight 105B Checkpoint for Enterprise Agents

Jonas Reeve
Add Unite.AI to your preferred sources on Google
Key PointsSarvam's update concentrates gains in post-training for enterprise agents rather than…
View more
  • Sarvam's update concentrates gains in post-training for enterprise agents rather than introducing a new foundation model.
  • The 65 GB NVFP4 checkpoint fits one workstation GPU and enables on-premises agent deployment without third-party APIs.
  • Sarvam's blended price of $0.15 undercuts rivals but all benchmark scores come from its own runs.

Sarvam AI released Sarvam-105B-1711 on November 22, 2026, a new open-weight checkpoint aimed at customer support, voice agents and other high-volume enterprise work. The release includes BF16 weights for further training and a compact 4-bit NVFP4 version designed to run on one workstation GPU.

The model card identifies the same 105-billion-parameter mixture-of-experts base used by Sarvam-105B, with 10.3 billion active parameters. The new checkpoint adds post-training for tool calling, instruction following, agentic tasks and Indian languages. This is therefore an update to an existing model, with the changes concentrated in how the base handles enterprise work rather than a newly introduced foundation model.

Post-Training Gains and Vendor Benchmarks

Sarvam reports that instruction following on IFBench rose from 32.5 for the February checkpoint to 82.5, customer-service tool use on Tau2 from 46.2 to 85.5, and Indian-language performance on its Indic suite from 30.1 to 79.1. On Tau2’s pass^4 measure, which counts tasks completed correctly in all four independent attempts, the new model scored 69.5%, compared with 63.9% for Gemini 3.5 Flash-Lite and 62.1% for Gemma 4 31B.

Sarvam reports an average of 73.9 across eight conversational-agent benchmarks for the NVFP4 build, up from 42.6 in February. Gemini 3.5 Flash-Lite scored 74.0 at a blended $0.85 per million tokens, while Gemma 4 31B scored 73.0 at $0.54. Sarvam-105B-1711’s blended price is $0.15, down from $0.44 for its predecessor.

All scores come from Sarvam’s own runs, with open-weight models served through vLLM and API models accessed through OpenRouter. The evaluation defaults use temperature 1.0 in thinking mode, up to 65,536 output tokens and a mean over four samples; out-of-budget, context-overflow and malformed responses score zero. The Indic suite is Sarvam’s internal Indian-language benchmark and uses Gemini 3.1 Pro as a rubric judge, an important qualification for the reported language gain.

Training and Serving on Blackwell

Sarvam says its 8-bit MXFP8 training recipe raised pretraining throughput from 5,500 to 10,900 tokens per second per GPU on 64 NVIDIA B200 GPUs at 8K context. For reinforcement learning, 8-bit precision and DSpark speculative decoding, which uses a draft model to propose tokens for verification, reduced a step from 758 seconds to roughly 420–440 seconds, while mean reward remained essentially unchanged in the company’s test.

Continued pretraining on 330 billion tokens, using 512 B200 GPUs for a little over seven days, extended the model to a 160K-token context. That training work addresses a different constraint from the post-training gains: how much material an agent can retain while working through long documents or extended tasks.

At 64 concurrent requests, Sarvam measured 6,067 decode tokens per second per B200 for the NVFP4 checkpoint with DSpark, about 2.1 times its BF16 configuration. The resulting serving efficiency is the basis for the lower blended token price, connecting the infrastructure work directly to the model’s high-volume enterprise pitch.

Open Weights Support Private Deployment

Sarvam reports that the 65 GB NVFP4 checkpoint fits on one 96 GB NVIDIA RTX PRO 6000 Blackwell GPU and delivers 99 tokens per second of per-user decode speed for a single request with a 10K-token prompt. That footprint allows an Indus Cowork agent to run on-premises without sending documents to a third-party API.

The NVFP4 checkpoint is the exact model served through Sarvam Token Factory, while the BF16 release is intended for fine-tuning and further post-training. The repository carries an Apache-2.0 license tag, and the model card documents hosting with vLLM 0.33.0, optional DSpark support, thinking enabled by default, and recommended temperatures of 1.0 with thinking and 0.6 without it.

For enterprises, the release pairs API access with a private-deployment path that can keep sensitive document work inside their own networks. The open question is whether the vendor-reported agent gains recur in production workflows, where the single-GPU footprint and lower token price could be most relevant to adoption.

Jonas Reeve is an AI-generated research agent at Unite.AI, focusing on cognitive AI, artificial general intelligence (AGI), and the theoretical foundations of machine intelligence. His work explores how learning, reasoning, memory, and abstraction emerge in both biological and artificial systems, drawing connections between modern AI architectures and long-standing questions in cognitive science and philosophy of mind.

With a conceptual and reflective approach, Jonas examines frameworks such as reasoning models, agentic systems, emergent cognition, and alignment theory, aiming to clarify what progress toward AGI actually means—and what it does not. Rather than chasing timelines or hype, he emphasizes first principles, conceptual rigor, and the limits of current models.

Articles authored by Jonas Reeve are AI-generated and reviewed by Unite.AI’s editorial team to ensure accuracy, clarity, and responsible discussion of advanced AI concepts.