TechPrufer
Get started
Cohere
Cohere logoCohere · Feature trail

Enterprise AI Translation Model

6 posts · Jul 2026 – Oct 2026

What this trail covers

Cohere's research into LLM training data reveals a 'culture funnel' where cultural diversity significantly narrows from pre-training to post-training datasets. This analysis highlights that scaling multilingual capabilities alone does not guarantee culturally rich representation. The findings emphasize the need for intentional curation of culturally representative content within training data to build more globally inclusive AI systems. This iteration introduces a framework for surfacing and examining cultural representation in training data and develops models capable of in-language reasoning across a wide range of languages, improving accessibility and nuance in AI interactions.

All Cohere trails
  1. In-Language Reasoning: Data Mixing Bridges | CohereThis post details the development of Tiny Aya L2-Thinker (3.35B), a massively multilingual model capable of in-language reasoning across 60 languages. The key technical contribution is a data mixing strategy comprising three pillars: English reasoning data (ER) for task-solving backbone, multilingual reasoning data (MR) for generalization, and multilingual non-reasoning data (NR) to boost accuracy and enable reasoning in low-resource languages. The model achieves over 93% in-language reasoning rate with minimal accuracy loss on most benchmarks, demonstrating improved cultural reasoning and accessibility.
  2. Introducing North Small Translate: A leading sovereign open-weight machine translation modelThis post introduces North Small Translate, a new mixture-of-experts machine translation model (218B total parameters, 25B active) with a 16k context length, supporting 50+ languages. It achieves an 83.6 WMT26 All Languages benchmark score, outperforming competitors like DeepL and Google Translate. The model demonstrates 1.4x higher output throughput than Gemma 4 31B and scores significantly higher on long-context evaluations. It is also cost-efficient, with a low per-task cost compared to other models. An agentic version, North Small Translate (Agentic), is also mentioned, scoring 84.36 on the WMT benchmark.
  3. Cohere Releases Command A+ for Sovereign AI | CohereCohere has released Command A+, a new open-source Mixture-of-Experts (MoE) model under the Apache 2.0 license. This model is purpose-built for sovereign critical infrastructure, offering transparency, control, and efficiency for enterprise and public sector organizations to deploy AI at scale with high security and data sovereignty. Key features include on-premises and private cloud deployment options (including air-gapped), full visibility into model architecture and behavior, and predictable costs without vendor lock-in. Command A+ is engineered for enterprise workloads such as RAG pipelines, SQL generation, and financial document analysis, with multimodal capabilities. Its MoE architecture (218B parameters total, 24B active per prompt) enables efficient inference on limited hardware (e.g., 2x H100s or 1x B200 GPU). It supports 48 languages, addressing global access and regulatory compliance needs like the EU AI Act.
  4. Tiny Aya Expedition Drives Multilingual Innovation | CohereThis post details the outcomes of the Expedition Tiny Aya initiative, showcasing projects that leveraged the Tiny Aya model. Key technical contributions include: the development of a multilingual Math Olympiad benchmark dataset for evaluating structured reasoning in math problem solving; a child-safe, offline AI voice assistant combined with a multilingual benchmark for evaluating child-focused AI systems; investigations into multilingual safety by testing behavior under code-mixed prompts and analyzing cross-lingual emergent misalignment; research into cultural conflicts in multilingual AI by examining model responses to manipulated language, geography, and context; a synthetic data generation pipeline for creating large-scale speech-to-speech translation datasets for under-resourced languages; an on-device multilingual retrieval and question answering system for document understanding; a parameter-efficient approach to connecting visual encoders to Tiny Aya for lightweight multilingual vision-language systems; and a system for on-device multilingual tool use by converting user requests into structured JSON tool calls, including optimized 4-bit quantized model variants.