How Chutes (SN64) Spent the Week Making Expensive GPUs Feel Awkward

Chutes (SN64) closed out last week with a dense shipment stack that keeps repositioning what decentralized inference can deliver.

Kimi K3, the largest open-weight model to date, went live one day after Moonshot dropped the weights. Parallax hit a training-speed milestone where consumer RTX 5090s outperformed professional RTX Pro 6000s.

DropZone Episode 1 arrived as the first public walkthrough of how the team pre-trains 20B models for under $10 per compute hour.

What Shipped During The Week

Four shipments carried the week, each moving Chutes’ position on frontier inference and decentralized training forward.

1. Kimi K3 live on Chutes day-one: 2.8T parameters, 1M-token context, native text/image/video input, $3 per million in and $15 per million out, running inside a hardware-attested TEE.

2. Parallax beats DDP on consumer cards: 8x RTX 5090s trained a 5B test model 12% faster than optimized DDP on 8x RTX Pro 6000s, at 65% lower hardware cost. On identical hardware, Parallax runs 47% faster than DDP.

3. DropZone Episode 1 out: Full walkthrough of the Parallax approach to sub-$10-per-hour 20B pre-training.

4. Nemotron 3 Nano Omni 30B live at the cheapest tier: $0.0245 per million input and $0.0978 per million output, 128K context, TEE by default.

Also Notable

Other noteworthy events during the week include:

1. Harvard research collaboration wrapped. Data collection ended July 24th.

2. Catalog housekeeping July 31st. GLM 5, Kimi K2.5, and Minimax M2.5 removed as low-usage.

3. Coming next: Hardware-attested registry access, independently verifiable TEE measurements, mutual TLS on all confidential VMs, and expanded Parallax pipeline coverage.