Skip to main content
Live
Main content

Sarvam releases Sarvam-105B-1711, an open-weight model that runs on one GPU

The mixture-of-experts checkpoint jumps from 42.6 to 73.9 on Sarvam's agent benchmark average and fits on a single 96 GB Nvidia card.

Jaeden Schafer
Editor in Chief · · 5 min read
Sarvam releases Sarvam-105B-1711, an open-weight model that runs on one GPU

Sarvam AI released Sarvam-105B-1711 on Nov. 22, a new open-weight checkpoint of its 105-billion-parameter model that lifts its eight-benchmark conversational-agent average from 42.6 to 73.9 and cuts its blended price from $0.44 to $0.15 per million tokens. The Bengaluru company is pitching the model for high-volume work such as customer support, voice agents and agentic tasks, and says its 4-bit version runs on a single Nvidia RTX PRO 6000 Blackwell card.

Sarvam-105B-1711 is a mixture-of-experts design with 105 billion total parameters and 10.3 billion active per token, according to its model card. Sarvam describes it as a new checkpoint of Sarvam 105B, building on the open-weight 30B and 105B models it released in March. Continued pretraining on 330 billion tokens, run across 512 Nvidia B200 GPUs for a little over seven days, extended the context window to 160,000 tokens.

The largest gains came in instruction following and tool use. The new checkpoint scored 82.5 on IFBench, up from 32.5, and 85.5 on the Tau2 agent benchmark, up from 46.2. Its score on Sarvam's own Indic-language suite rose from 30.1 to 79.1, and a voice fine-tune scored 68.7 on the company's Samvaad benchmark, which tests next-turn responses across English and 12 Indian languages, up from 57.9 for the base model.

Key facts

  • 01Sarvam-105B-1711 lifts Sarvam's eight-benchmark conversational-agent average from 42.6 to 73.9.
  • 02The mixture-of-experts model has 105 billion total parameters and 10.3 billion active, per its model card.
  • 03Blended API price falls from $0.44 to $0.15 per million tokens on Sarvam's Token Factory.
  • 04The 65 GB NVFP4 build runs on a single 96 GB Nvidia RTX PRO 6000 Blackwell at 99 tokens per second per user.
  • 05Context extends to 160,000 tokens after continued pretraining on 330 billion tokens across 512 B200 GPUs.

Sarvam places the model in the Flash-Lite class and published head-to-head numbers against Gemini 3.5 Flash-Lite, Gemma 4 31B, GPT-6 Luna, Qwen3.8 27B, Nemotron 3 Super and Mistral Small 4. On the eight-benchmark average, Sarvam's 73.9 sits just behind Gemini 3.5 Flash-Lite at 74.0 and Qwen3.8 27B at 74.6, and ahead of Gemma 4 31B at 73.0 and GPT-6 Luna at 66.3. Sarvam says it leads the group on IFEval, IFBench and Arena-Hard v2 hard prompts and lands within 0.3 points of the best Tau2 average.

Sarvam's own table also shows where the model trails. On Humanity's Last Exam without tools it scored 12.3, against 30.6 for GPT-6 Luna and 30.8 for Qwen3.8 27B, and on Arena-Hard v2 creative-writing prompts it scored 65.4 to Gemini 3.5 Flash-Lite's 90.2. On Tau2 pass^4, a reliability measure that requires an agent to succeed on all four attempts at a task, Sarvam reported 69.5, the top figure in its comparison.

Price is the sharper pitch. Sarvam's Token Factory API lists the model at 10 rupees per million input tokens and 25 rupees per million output tokens, excluding GST, which the company converts to a blended $0.15 per million tokens at a 3-to-1 input-to-output ratio. On the same basis, Sarvam puts Gemini 3.5 Flash-Lite at $0.85, Gemma 4 31B at $0.54, GPT-6 Luna at $0.20 and Nemotron 3 Super at $0.17.

The NVFP4 build weighs about 65 GB and fits on one 96 GB RTX PRO 6000 Blackwell, where Sarvam measured 99 tokens per second for a single user with a 10,000-token prompt using vLLM and speculative decoding. On a B200 serving 64 concurrent users, the company reported about 6,000 tokens per second per GPU, which it said is twice its previous setup. Sarvam is targeting on-premises deployments for banks, government agencies and other strategic sectors.

Sarvam also reported training-side gains. Moving pretraining to 8-bit MXFP8 roughly doubled throughput on 64 B200s, from 5,500 to 10,900 tokens per second per GPU, and reinforcement-learning step time fell from about 13 minutes to about seven with no drop in reward, according to the company.

Related · from this week
Mistral releases open weights for Large 4, its 1-trillion-parameter model
Jaeden Schafer · 5 min read →

The 4-bit weights are available on Hugging Face under sarvamai/sarvam-105b-1711, with a separate BF16 version for teams that want to train further. The NVFP4 version is the one Sarvam serves through Token Factory, and the company has also published its Samvaad voice benchmark as a Hugging Face dataset.

ShareXLinkedInEmail
AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

  • ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat
  • Generate images & video with Sora, Veo, Ideogram
  • Compare any two models side by side
  • From $8.99/mo · 80+ models, all included
Try AI Boxaibox.ai
Trusted by 3,000+ teams
Got a tip?

Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.

Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.

AI Box Daily briefingFree · Daily · No fluff

Stay ahead of everyone in AI.

The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.

Loved by 10,000+ AI professionals
Free forever. Unsubscribe with one click.

The briefing read inside teams at

Keep reading

More from Models

Mistral releases open weights for Large 4, its 1-trillion-parameter model
Models

Mistral releases open weights for Large 4, its 1-trillion-parameter model

The checkpoint lands on Hugging Face three weeks after preview, scoring 41% on Terminal-Bench 4.0 and priced at $0.70 per million input tokens via API.

Jaeden Schafer5 min read
Nvidia logo
Models

Nvidia releases Nemotron 3.5 Transcribe, an open speech model for 40 languages

The promptable model uses about 4 billion active parameters, ranks first among open models on Hugging Face's English ASR leaderboard and runs on a single GPU.

Jaeden Schafer4 min read
Nvidia logo
Models

Nvidia opens Alpamayo 2 Super for commercial robotaxi deployment

The 30B-parameter driving model tops LingoQA against Qwen, Gemini and GPT-4o, and ships under a permissive Linux Foundation license.

Jaeden Schafer5 min read