Now in Foundry: Cohere Transcribe, Nanbeige 4.1-3B, and Octen Embedding

2026-04-06 · Source: Microsoft Foundry Blog articles · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Cloud Computing & IT Infrastructure · Depth: Intermediate, medium

Summary

This week's Model Mondays highlights three open-source AI models available through Microsoft Foundry, each addressing a different layer of the AI application stack. Cohere's cohere-transcribe-03-2026 is a 2B ASR model that leads the Open ASR Leaderboard with a 5.42% average Word Error Rate across 8 English benchmarks and supports 14 languages with automatic long-form audio chunking. Nanbeige's Nanbeige4.1-3B, a 3B reasoning model, surpasses larger models in coding, math, and deep-search benchmarks, scoring 76.9 on LiveCodeBench-V6 and 73.2 on Arena-Hard-v2, and offers native tool-use support. Octen's Octen-Embedding-0.6B, a 0.6B text embedding model, achieves a 0.7241 mean task score on the RTEB leaderboard, outperforming larger proprietary models, and provides 32,768-token context for long-document retrieval across 100+ languages and specialized domains.

Key takeaway

For AI Architects and NLP Engineers building production-grade systems, these models demonstrate that high performance across speech, reasoning, and retrieval can be achieved with compact, open-source solutions. You should consider integrating these specialized models via Microsoft Foundry to optimize for cost and efficiency without sacrificing accuracy, especially for multilingual or domain-specific applications. Evaluate their benchmarks against your specific requirements before committing to larger, more resource-intensive alternatives.

Key insights

Specialized, smaller open-source models can outperform larger general-purpose models across various AI tasks.

Principles

• Targeted fine-tuning improves performance
• Dedicated architectures enhance efficiency
• Preference alignment boosts output quality

Method

Deploy open-source models from Hugging Face directly via Microsoft Foundry's catalog or "Deploy on Microsoft Foundry" option from the Hugging Face Hub for secure, scalable inference.

In practice

• Use cohere-transcribe for multilingual meeting transcription
• Apply Nanbeige4.1-3B for complex agentic workflows
• Employ Octen-Embedding-0.6B for domain-specific semantic search

Topics

• Cohere Transcribe
• Nanbeige 4.1-3B
• Octen Embedding
• Automatic Speech Recognition
• Text Embeddings

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

Related on AIssential

• Introducing MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 in Microsoft Foundry — Microsoft Foundry introduces three new AI models for efficient, high-quality speech recognition, speech generation, and text-to-image capabilities.
• New Open Audio Models 🤗 | Recap with Jeff — New open audio models from Mistral and Cohear offer efficient text-to-speech and speech-to-text capabilities.
• Solving the Embedding Step in NLP Projects — Effective embeddings are crucial for NLP model generalization and reliable downstream task performance.
• A New Chapter for Realtime AI: Reasoning, Translation, and Real-Time Transcription — New OpenAI models enhance real-time AI voice applications with advanced translation, transcription, and reasoning capabilities.
• Mistral drops new speech-to-text AI models — Mistral's new speech-to-text models prioritize speed, privacy, and affordability for diverse enterprise applications.
• Cohere cracks lossless quantization and native citations with first full Apache 2.0 licensed open model Command A+ — Cohere's Command A+ combines sparse MoE architecture with lossless 4-bit quantization for efficient, powerful enterprise AI.

Open in AIssential →

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