Marin's 535B run passes 80% of pretraining, beating its scaling-ladder forecast
dlwh · x · 2026-10-24
Marin's 535B model is now over 80% through pretraining. The loss-vs-scaling-ladder forecast chart shows training beating the prediction — but the author notes this isn't ideal: "we probably messed up the asymptote" and will refine the forecasts. The sharp loss drop comes from the planned 4K→16K context switch, which effectively cheats the forecast; the team will re-evaluate at 4K as well.
More from Infra
- RTX 30-series GPU Unlock tested: up to 59% throughput gain in ComfyUI on 3090 Ti — under-_-null · 2026-10-24
- RTX 3070 runs quantized Qwen3.8 at 1000 tps prefill with Strata — full config shared — SetNo7692 · 2026-10-24
- OpenAI resumed frontier RL training on Oct 1, 11 days after the containment pause — scaling01 · 2026-10-24
- Cloudflare cuts CPU use 30% on D1 backend with new profiler, Workers Observability drops memory/CPU ~50% — irvinebroque · 2026-10-24
- NVIDIA makes 8-node DGX Spark clusters official: 97.4% scaling efficiency on 900M training — TheZachMueller · 2026-10-24
- One NVIDIA Vera Rubin rack would set you back at least ~$7M — LinusEkenstam · 2026-10-24