Sarvam reduces the cost and size of its model for enterprise agents
Sarvam AI has released Sarvam-105B-1711, an open weight checkpoint aimed at enterprise agents, customer support, and voice. The model fits on a 96 GB workstation GPU and targets local deployment without third-party APIs.
Sarvam AI released Sarvam-105B-1711 on November 22, 2026, an open weight checkpoint designed for enterprise agents, customer support, and high-volume voice use cases. The model does not replace a new foundation: it builds on the 105 billion parameter Mixture-of-Experts base of Sarvam-105B, with 10.3 billion active parameters, then adds post-training for tool calling, instruction following, agentic tasks, and Indian languages. The concrete goal is to run these workflows on-site, with a lower announced cost and without relying on third-party APIs.
A model sharpened for execution, not for show
The Sarvam update focuses on the post-training phase, where a model learns to better follow instructions and use tools. Two versions are highlighted: BF16 weights for continued training, and a 4-bit NVFP4 variant designed to run on a single workstation GPU. In practice, Sarvam aims less to impress than to make inference easier to deploy. The company thus presents its checkpoint as a tool for agents who must respond, call tools, and process documents without running a small power plant in the server corner.
In its own tests, Sarvam says it raised the IFBench instruction-following score from 32.5 to 82.5 between the February checkpoint and this version, while Tau2 for tool use in customer service rose from 46.2 to 85.5. The Indic suite, its internal benchmark for Indian languages, improved from 30.1 to 79.1. At this stage, these figures remain in-house results: evaluations use vLLM for open weight models, OpenRouter for API models, and a protocol where out-of-budget or malformed responses count as zero. The increase is clear, but it therefore requires [to be verified] on real workloads.
What the NVFP4 format changes for enterprises
The NVFP4 format is the core of the product message. Sarvam indicates that the 65 GB checkpoint fits on an NVIDIA RTX PRO 6000 Blackwell 96 GB GPU and reaches 99 tokens per second per user with a 10,000 token prompt. This allows, according to the company, deploying an Indus Cowork agent internally, without sending documents to an external service. For enterprises, the issue goes beyond performance alone: it touches confidentiality, latency, and cost control. Teams handling sensitive data primarily look at this triptych, not the logo on the homepage.
The model served via Sarvam Token Factory is precisely this NVFP4 version, while the BF16 version is intended for fine-tuning and additional post-training. The repository is announced under Apache-2.0 license, with a service via vLLM 0.33.0 and, optionally, DSpark. Sarvam also recommends a “thinking” mode enabled by default, with a temperature of 1.0 in this mode and 0.6 without it. In other words, the company pushes a ready-to-use model but keeps the door open for internal adaptations.
Infrastructure gains that also benefit the price
Sarvam’s pitch is not only about the model but about the execution chain. The company claims its MXFP8 training recipe on 64 NVIDIA B200 GPUs increased pretraining throughput from 5,500 to 10,900 tokens per second per GPU, at 8K context. For reinforcement learning, the use of 8-bit precision and DSpark speculative decoding reportedly reduced a step from 758 seconds to about 420 to 440 seconds, without notable drop in average reward in its tests. The extended pretraining, on 330 billion tokens with 512 B200 GPUs over just over seven days, expanded the context to 160K tokens.
These technical choices explain the displayed price. Sarvam announces a blended cost of 0.15 dollars per million tokens for Sarvam-105B-1711, versus 0.44 dollars for its predecessor. In its own comparisons, Gemini 3.5 Flash-Lite comes out at 0.85 dollars per million tokens for an average score of 74.0 on eight conversational agent benchmarks, while Gemma 4 31B scores 73.0 at 0.54 dollars. Sarvam claims 73.9 at 0.15 dollars. The picture is appealing, but one detail matters: all scores come from the publisher’s runs.
Why this launch matters for enterprise agents
The movement is clear: Sarvam pushes an open weight model toward use cases where the agent must be hosted locally, handle sensitive documents, and remain economical at scale. This combination interests business tool publishers, integrators, and large organizations that do not want to depend on an external API for every request. The question is therefore not only raw performance but the ability to maintain these gains in production, on real workloads and business constraints tougher than those of a benchmark.
In this logic, Sarvam-105B-1711 looks less like a lab demonstration and more like another piece in the competition around deployable agent models, efficient in computation and compatible with private environments. It remains to be seen if the gains measured internally hold up outside the company’s controlled setting. That’s where the story gets interesting, and a bit less docile.
Key points
- November 22, 2026: release of Sarvam-105B-1711.
- 105 billion parameters, including 10.3 billion active.
- 65 GB for the NVFP4 version, according to Sarvam.
- 0.15 dollars per million tokens, announced blended price.
- Sarvam cites Gemini 3.5 Flash-Lite and Gemma 4 31B.
By the numbers
- 5,500 to 10,900 tokens/s/GPU — pretraining throughput on 64 NVIDIA B200, according to Sarvam.
- 330 billion tokens — continuation of pretraining on 512 B200, according to Sarvam.
- 160K tokens — extended context after continuous training, according to Sarvam.
- 99 tokens/s — announced throughput for one user on RTX PRO 6000 Blackwell 96 GB.
- 69.5% — Tau2 pass^4 score claimed by Sarvam in its tests.