DeepSeek rerun through Fireworks reproduces results; original dip was likely a fluke
teortaxesTex · x · 2026-10-28
After their initial run showed the new DeepSeek version slightly worse than its predecessor, the team delayed publication to investigate, suspecting issues with inference provider Fireworks. Following discussions with Fireworks, a rerun now reproduces their results: no underlying issue was found to explain the earlier significantly worse performance, which appears to have been a temporary fluke rather than a model regression.
More from Infra
- Machine traffic passes 50% on Cloudflare, but its bot-paywall has earned nothing yet — FinanceYF5 · 2026-10-29
- Citadel Securities: token prices down ~50% in six months, frontier lab margins getting crushed — GaryMarcus · 2026-10-29
- Meta aggregates stranded idle GPU capacity into an opportunistic pool for ads ranking — Meta_Engineers · 2026-10-29
- LLM token prices crash 55% from May peak, now below where the year started — GaryMarcus · 2026-10-29
- NVIDIA Details Serving Cosmos 3 Super World Models on Nebius: Latency vs Throughput — NVIDIAAI · 2026-10-29
- CXMT runs 66% operating margins as China's memory chip hill-climb strategy pays off — PAstynome · 2026-10-28