Cerebras Running Kimi K2.6

A bundle of Cerebras-related X (Twitter) threads is included, with the recent one centered on a performance boost for Cerebras Inference. The company stated that it increased Llama3.1-70B processing speed from 650 t/s to 2,100 t/s, claiming it is 16x faster than GPUs.

In earlier announcements, Cerebras Inference was described as delivering 1,800 tokens/s on 8B models and 450 tokens/s on 70B models, highlighting the speed and price competitiveness of the Llama3.1 inference API. It also noted that first-token response latency matters for real-time applications, emphasizing the advantages of wafer-scale integration.

The thread also includes other past announcements.

• BTLM-3B-8K: introduced as a 3B-parameter model aiming for 7B-class performance with an 8K context
• Condor Galaxy-1: announcement of a 4 exaflop-class AI supercomputer built from 64 CS-2 systems
• SlimPajama-627B: a large-scale public dataset created by cleaning and deduplicating RedPajama
• Cerebras-GPT: release of GPT-series models ranging from 111M to 13B in scale

Overall, this shows Cerebras simultaneously pushing forward on inference performance, large-scale training infrastructure, and open models/datasets.