Kimi K3.1 Pro is live with open weights on Nov 4. Kernel module doubles BF16 throughput on RTX 30-series. Cerebras serves 3-trillion-parameter models at over 1,000 tokens per second. Harness pushes open models past Fable 5 at about 9% of the cost.
Key points
Kimi K3.1 Pro live, open weights Nov 4
Kernel module doubles BF16 on RTX 30-series
Cerebras serving 3T models at 1,000+ tok/s
Harness pushes open models past Fable 5 at ~9% the cost
Sources
Neolab neocloud pivot locks GPU capacity, prices up 2–3x
Neolab neocloud pivot locks GPU capacity early and notes H100/B200/B300 rental prices have risen 2–3x. They intend to resell the compute to generate revenue.
Key points
GPU capacity: locked in early.
Rental prices: H100/B200/B300 up 2–3x.