Cohere's Aya Expanse 8B beats models twice its size by pitting teacher AIs against each other

Cohere For AI just dropped the Aya Expanse family, and the 32B model is punching so far above its weight class that it’s beating Llama 3.1 70B in multilingual benchmarks. Let that sink in. A model with less than half the parameters is winning in pairwise comparisons across 23 languages, including on the Arena-Hard-Auto dataset that Cohere translated and is releasing publicly. The 8B version isn’t slacking either—it’s clocking win rates between 60.4% and 70.6% against Gemma 2 9B, Llama 3.1 8B, and Mistral’s Ministral 8B.

What makes this work isn’t just scale. It’s method. The team tackled two persistent headaches in multilingual AI: the lack of strong teacher models for low-resource languages and the fact that preference training tends to cause regressions when you optimize for too many languages at once. Their answer is what they call “data arbitrage.” Instead of relying on a single teacher model to generate synthetic training data, they train a pool of models on different language groups and use an internal reward model as an arbiter. For each prompt, every model in the pool generates a completion, the arbiter scores them, and only the best one makes the cut. The result? Even at the supervised fine-tuning stage, the 8B model saw a 9.1% win-rate improvement against Gemma 2 9B compared to Cohere’s previous Aya 23 model.

This flips a common assumption on its head. The industry has been racing toward monolithic, ever-larger teacher models, but Cohere’s arbitrage approach suggests that a diverse committee of specialized models can produce better multilingual data than any single giant. It’s a practical insight for teams that don’t have the compute to train a GPT-4-class teacher across 100 languages. You don’t need one perfect model—you need a smart way to pick the best output from several imperfect ones.

The preference optimization pipeline is equally clever. Most multilingual preference data is either English-only or low quality, so Cohere synthesized their own by contrasting strong in-language completions with weaker ones translated from English. This sidesteps the optimization nightmare where improving Hindi performance accidentally nukes French. The models are open-weight, which means researchers can actually pick apart what’s happening under the hood rather than taking a vendor’s word for it. For anyone building products that need to work across languages without the luxury of per-language fine-tuning, Aya Expanse just raised the bar and lowered the entry cost.