Sarvam releases upgraded 105B AI model with cheaper hosting and a single-GPU option
Sarvam-105B-1711 pairs downloadable weights with a $0.15 blended price per million tokens. Its claimed capability gains come from company-run tests.
Loading page…
Sarvam-105B-1711 pairs downloadable weights with a $0.15 blended price per million tokens. Its claimed capability gains come from company-run tests.
Listen to this story
Sarvam’s November 22 update makes its 105B checkpoint more viable for hosted agent workloads and controlled on-prem deployments: serving is cheaper, and a compressed weight set is designed for one workstation-class GPU. Sarvam reports near-parity with Gemini 3.5 Flash-Lite on an eight-benchmark average, but results vary by task and come from company-run evaluations, so they do not establish broad interchangeability. Teams should also distinguish the downloadable checkpoint from the voice-specific fine-tune used in Samvaad.
Sarvam lists blended hosted inference at $0.15 per million tokens, down from $0.44, based on three input tokens per output token; this is not an estimate of self-hosting costs.
The 65GB NVFP4 download is intended to run on a single Nvidia RTX PRO 6000 Blackwell workstation card with 96GB of memory, not typical consumer hardware.
Sarvam reports an eight-benchmark average of 73.9 versus Gemini 3.5 Flash-Lite’s 74.0, at listed blended prices of $0.15 and $0.85 per million tokens, respectively.
Businesses running customer-support and voice agents have a cheaper hosted option from Sarvam, plus downloadable weights the company says fit one workstation graphics card. Released November 22, 2026, Sarvam-105B-1711 updates its existing 105-billion-parameter model, pairing a lower serving price with company-reported gains in following instructions, using tools and handling Indian languages.
Sarvam lists a blended inference price—the charge for processing inputs and generating outputs—of $0.15 per million tokens, down from $0.44. That is a hosted-service price, not the cost of operating the downloadable model on a customer’s own equipment. The calculation assumes three input tokens for every output token; it should not be read as a flat charge for a million output tokens.
The private-deployment option centers on a compressed, 65GB download. Sarvam says it runs on a single Nvidia RTX PRO 6000 Blackwell graphics card with 96GB of memory. That qualification matters: the claim concerns a workstation card, not an ordinary laptop or consumer PC. Sarvam is pitching it to banks, government departments and other organizations that want agents running on their own premises.
The weights—the model’s learned values—are available on Hugging Face in two numerical formats. These are versions of the same checkpoint, rather than separate foundation-model launches:
Sarvam reports an average score of 73.9 across eight conversational-agent benchmarks, up from 42.6 for the earlier checkpoint. The group covers instruction following, conversation quality, tool calling and Indian languages. In the same evaluation, Gemini 3.5 Flash-Lite averaged 74.0 at a listed blended price of $0.85 per million tokens. That is the basis for Sarvam’s claim of comparable performance at less than one-fifth the price—not a finding that the models are interchangeable across all work.
The company’s table shows why the task matters. Sarvam scored 85.5 on Tau2, which tests tool use in customer-service conversations, against Gemini’s 83.1. But it trailed Gemini on BFCL v4’s broader tool-calling measure, 62.4 to 67.8. Sarvam also reports that 69.5% of Tau2 tasks were completed correctly in all four attempts, compared with 63.9% for Gemini—a separate measure of consistency rather than average success.
These are company-run evaluations, not independent deployment results. Its internal Indic-language suite rose from 30.1 to 79.1 and uses a language model to judge answers. The higher-precision and compressed downloads also do not score identically on every task, so the model format matters when reading a comparison.
Sarvam attributes the lower serving cost to changes on Nvidia Blackwell hardware. It says its four-bit model generates about 6,000 tokens per second on one B200 GPU across 64 concurrent users, twice its previous setup. That is shared throughput across those users, distinct from the workstation claim. Sarvam says the serving improvements roughly halve its cost per token.
Training changes extended the context window—the amount of material the model can work with at once—to 160,000 tokens. Sarvam says that stage processed 330 billion tokens across 512 GPUs in a little over seven days. Separately, it reports that eight-bit training increased pretraining throughput from 5,500 to 10,900 tokens per second per GPU, allowing the same cluster to process roughly twice as many tokens.
The checkpoint is also feeding Sarvam’s applications, but not always unchanged. Its Samvaad voice-agent service uses a voice-specific fine-tune of Sarvam-105B-1711. That distinction gives businesses evaluating the download a practical boundary: the published checkpoint and the specialized model inside the voice service are not the same offering.
Loading discussion...
Join the conversation
Explain which work you would keep inside the organization.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.