Skip to main content
Back to Pulse
TechCrunch+1 source

Cohere launches a family of open multilingual models

Read the full articleCohere launches a family of open multilingual models on TechCrunch

What Happened

Cohere's Tiny Aya models support over 70 langauges

Our Take

Cohere's right to go open and multilingual, but "over 70 languages" is a press release, not a product. The question's whether Aya actually competes on quality or just breadth.

Multilingual models are getting commoditized fast. Open + fine-tuning = you can support any language if the base is good enough. Cohere's betting they're good enough (probably they're fine—not amazing, not trash). Good for specific cases: customer support in Hindi, moderation in Turkish, etc.

But "open model for many languages" won't differentiate if base capability's average. We'll know in three months when people ship with it.

What To Do

If you need multilingual inference cheap, test Aya against Claude Opus on your specific languages—probably saves money, might lose quality.

Builder's Brief

Who

teams building multilingual NLP pipelines for non-English markets

What changes

open weights eliminate per-token API costs for small-model multilingual inference on-device or self-hosted

When

weeks

Watch for

community benchmark scores on low-resource languages (Swahili, Bengali, Yoruba) compared to mBERT and Llama multilingual variants

What Skeptics Say

70-language support degrades sharply in low-resource languages where training data is thin; Cohere's open model ecosystem lags Meta and Alibaba in tooling, adoption, and community momentum, limiting real-world deployment.

Cited By

React

Newsletter

Get the weekly AI digest

The stories that matter, with a builder's perspective. Every Thursday.

Loading comments...