Developer burns $4,600 on rented GPUs with zero training steps, finishes same RL run on Tinker for under $100

simonguozirui · x · 2026-11-08

Developer msuiche shared a detailed cost comparison of fine-tuning DeepSeek V4.1 via Thinking Machines' Tinker platform:

Previous Modal attempt: $4,600 wasted on placement and clustered networking failures, zero training steps completed — manual retry loops never produced a comparable RLVR baseline
On Tinker: 60-step training leg $22 (token-metered), no queue/boot time; 48-episode eval sweep $13; the full campaign came in under $100
Self-managed baseline: 8×H200 at $28.8/hr ($197 for training) + $60–100 evals, plus 5–6 boot cycles lost to infra bugs — the equivalent campaign would cost $12,000

Tinker's team amplified the post, saying this is exactly the goal: cheap and easy enough that there's no reason to run it yourself. Tinker ships a working sampler (tokens + logprobs) and a native DeepSeek renderer, turning an expensive dead end into a measured arm in SwishBench-12.

Original post →

Developer burns $4,600 on rented GPUs with zero training steps, finishes same RL run on Tinker for under $100

More from Infra