Now in Foundry: NVIDIA Nemotron-3-Super-120B-A12B, IBM Granite-4.0-1b-Speech, and Sarvam-105B

2026-03-30 · Source: Microsoft Foundry Blog articles · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing, Speech Recognition · Depth: Advanced, medium

Summary

Microsoft Foundry's Hugging Face collection now includes three new models: NVIDIA Nemotron-3-Super-120B-A12B, IBM Granite-4.0-1b-Speech, and Sarvam-105B. NVIDIA's Nemotron-3-Super-120B-A12B is a hybrid Latent Mixture-of-Experts (MoE) model with 12B active parameters, supporting up to 1 million tokens and featuring configurable reasoning and native speculative decoding. IBM Granite-4.0-1b-Speech is a compact ~1B parameter ASR/AST model achieving a 5.52% average Word Error Rate (WER) at 280× real-time speed, with runtime keyword biasing and bidirectional translation for six languages. Sarvam-105B is a 105B MoE model with 10.3B active parameters, optimized for 22 Indian languages and English, demonstrating strong agentic performance on web search and task-planning benchmarks.

Key takeaway

For AI Architects evaluating large language models for specialized applications, these new additions to Microsoft Foundry offer distinct advantages. Consider NVIDIA Nemotron-3-Super-120B-A12B for ultra-long context and agentic workflows requiring configurable reasoning. IBM Granite-4.0-1b-Speech is ideal for compact, high-speed multilingual ASR/AST with dynamic domain adaptation. Sarvam-105B provides robust agentic capabilities and broad Indian language support, crucial for diverse global deployments.

Key insights

New models in Microsoft Foundry offer specialized capabilities for long-context, multilingual speech, and agentic tasks.

Principles

• Hybrid MoE architectures improve accuracy per parameter.
• Runtime keyword biasing enables domain adaptation without fine-tuning.
• Multi-token prediction reduces time-to-first-token.

Method

NVIDIA's Latent MoE architecture combines Mamba-2 state-space layers and sparse MoE layers with full attention, routing tokens to a smaller latent space for computation.

In practice

• Use Nemotron-3-Super for 1M-token RAG and code analysis.
• Employ Granite-4.0-1b-Speech for domain-specific ASR via keyword biasing.
• Leverage Sarvam-105B for agentic workflows in 22 Indian languages.

Topics

• NVIDIA Nemotron-3-Super-120B-A12B
• IBM Granite-4.0-1b-Speech
• Sarvam-105B
• Mixture-of-Experts
• Automatic Speech Recognition

Best for: AI Architect, AI Engineer, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Now in Foundry: IBM Granite 4.1, NVIDIA Nemotron Nano Omni, and Qwen3.6-35B-A3B — New models in Microsoft Foundry offer diverse capabilities from multimodal reasoning to agentic coding and document intelligence.
• NVIDIA’s New AI Just Changed Everything — NVIDIA's Nemotron 3 Super offers a powerful, open-source AI assistant with significant speed improvements via novel architectural and mathematical techniques.
• How NVIDIA Extreme Hardware-Software Co-Design Delivered a Large Inference Boost for Sarvam AI’s Sovereign Models — Full-stack co-optimization of model design, kernels, scheduling, and hardware significantly boosts LLM inference performance.
• Introducing the IBM Granite 4.1 family of models — IBM's Granite 4.1 models offer modular, efficient, and governable AI solutions for enterprise workflows, prioritizing data quality over raw quantity.
• NVIDIA Nemotron 2 Nano 9B Japanese: 日本のソブリンAIを支える最先端小規模言語モデル — NVIDIA's Nemotron-Nano-9B-v2-Japanese offers SOTA Japanese AI capabilities in a compact, deployable SLM.
• Mistral Small 4: A Good Alternative to Qwen3.5 122B and Nemotron 3 Super? — Mistral Small 4 unifies instruct and reasoning, while NVIDIA's new Nemotron models offer high performance with efficient KV cache.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Microsoft Foundry Blog articles.