Latest open artifacts (#20): New orgs! New types of models! With Nemotron Super, Sarvam, Cohere Transcribe, & others

2026-03-30 · Source: Interconnects AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, short

Summary

The latest Artifacts Log highlights a diverse array of new open models across various modalities and use cases, departing from the usual dominance of large models from major players. This month's roundup features models for optical character recognition (OCR), RAG search, audio transcription, computer-use, code-editing, and math theorem proving. Notable releases include NVIDIA's Nemotron-3-Super-120B-A12B-NVFP4, a 120B parameter model with a 1M context window and LatentMoE architecture; Cohere's cohere-transcribe-03-2026, a 14-language speech-to-text model under Apache 2.0; and Sarvam's sarvam-105b, an Indic-language optimized model from an Indian startup. Other models include Mistral-Small-4-119B-2603 for hybrid reasoning, Zed Industries' zeta-2 for code editing, and several multimodal and special-purpose models from IBM, Microsoft, and others, underscoring a growing trend towards domain-specific, cost-effective open models.

Key takeaway

For NLP Engineers and CTOs evaluating open-source model adoption, this log indicates a strong trend towards specialized, efficient models. You should prioritize exploring domain-specific models like Sarvam for regional language support or NVIDIA's Nemotron for specific architectural advantages, rather than solely focusing on general-purpose giants. This shift enables more targeted and potentially cost-effective solutions for your specific application needs.

Key insights

The open model ecosystem is diversifying with many specialized, smaller models complementing large, general-purpose ones.

Principles

• Domain-specific models are crucial tools.
• Sovereign AI drives language-specific model preference.
• Open data collection enhances model training.

Method

NVIDIA's Nemotron-3-Super uses LatentMoE and NVFP4 during pre-training. Yuan3.0-Ultra prunes experts after pre-training a 1.5T model on 2.2T tokens. Chroma's Context-1 is fine-tuned for agentic search using Thinking Machine's Tinker.

In practice

• Explore Nemotron-3-Super for large context window tasks.
• Utilize Cohere-transcribe-03-2026 for multi-language speech-to-text.
• Consider Sarvam models for Indic language applications.

Topics

• Open Model Ecosystem
• NVIDIA Nemotron-3-Super
• Cohere Transcribe
• Sarvam AI
• Multimodal AI

Code references

• Yuan-lab-LLM/Yuan3.0-Ultra

Best for: NLP Engineer, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Why Nvidia builds open models with Bryan Catanzaro — NVIDIA's Nemotron project strategically advances AI infrastructure through open models, data, and research, driven by both internal product needs and ecosystem growth.
• OpenAI's superapp hiding inside Codex — OpenAI is evolving Codex into a "superapp" with expanded agentic capabilities, while Anthropic's Opus 4.7 advances coding performance.
• Now in Foundry: NVIDIA Nemotron-3-Super-120B-A12B, IBM Granite-4.0-1b-Speech, and Sarvam-105B — New models in Microsoft Foundry offer specialized capabilities for long-context, multilingual speech, and agentic tasks.
• 2 NEW Secret Models 👀 just dropped! — New "secret" LLMs on OpenRouter, Hunter Alpha and Healer Alpha, offer advanced agentic and multimodal capabilities for free.
• NVIDIA's New Free AI - A Gift To Humanity — NVIDIA's NeMoTron 3 Ultra is a fast, open-licensed, text-only 550B parameter model excelling in agentic tasks but not complex coding.
• Latest open artifacts (#17): NVIDIA, Arcee, Minimax, DeepSeek, Z.ai and others close an eventful year on a high note — The open model ecosystem is rapidly evolving with diverse architectures, data transparency, and specialized performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Interconnects AI.