Chutes says it trained a 20B model for under $10 an hour using rented GPUs across two continents
markjeffrey · x · 2026-07-27
Chutes trained a 20B model for under $10/hour without a dedicated cluster
An OpenTensor ecosystem highlight says @chutesai trained a 20B model for under $10/hour by coordinating rented GPUs across two continents, without using a dedicated cluster.
The post frames it as a practical sign that distributed compute rentals can materially lower training costs for serious model work.
More from Infra
- Users ask whether RAID 0 NVMe setups improve local large-model runs on Pulsar — Wyldkard79 · 2026-07-27
- GLM-5.2 inference on RTX 5090s jumps from 30 tok/s to 80–110 tok/s — markjeffrey · 2026-07-27
- Hermes Agent says progressive tool disclosure scales MCP tools with near-zero accuracy loss — Teknium · 2026-07-27
- Laguna tests 2.75 and 3.25 bpw quantization with NVFP4 experts and FP8 KV cache — QuixiAI · 2026-07-27
- User reports 44 tok/s ingestion and 8 tok/s generation for GLM 5.2 on a $900 rig — naunen · 2026-07-27
- llama.cpp merges support for Minimax M3 with MSA — Time_Reaper · 2026-07-27