Xiaomi MiMo Achieves 15X Faster AI Inference Than ChatGPT and Claude

Xiaomi recently introduced its MiMo-V2.5-Pro-UltraSpeed model, which has set a new benchmark for artificial intelligence inference speed. This model can reportedly achieve over 1,000 tokens per second, peaking near 1,200 in demonstrations. Importantly, it accomplishes this feat using a standard 8-GPU commodity node, rather than specialized chips.

The speed improvement stems from various technical innovations, including FP4 quantization in the model’s expert layers and DFlash speculative decoding, which allows the system to propose an entire block of tokens during processing. This contrasts with many existing models that analyze tokens one at a time.

For context, the previous leaders in AI processing speeds, such as GPT-5.5 and Claude Opus 4.6, operated at much lower rates of around 68 to 71 tokens per second, while the Gemini Flash managed 192 tokens per second. Xiaomi's MiMo model significantly outpaces these competitors, even rivaling other high-efficiency models in coding benchmarks.

The implications of this innovation are substantial; it lowers barriers for organizations seeking to deploy high-speed AI capabilities without needing custom hardware solutions. As Xiaomi notes, "Our goal was to make this technology accessible without having to rely on expensive custom chips," which could reshape AI deployment strategies across various sectors.

A limited API trial for MiMo is expected to commence soon, providing users the opportunity to experiment with this technology. The pricing for this trial will be at three times the standard MiMo rates but with an estimated tenfold increase in output speed, further emphasizing the model's operational efficiency.

As the landscape of artificial intelligence continues to evolve, Xiaomi’s breakthrough with MiMo raises key questions about the future of AI capabilities in everyday applications and for discerning who can afford such performance in practice.