Reflection Ships Apache 2.0 Model With Tech Report; Analyst Estimates Pre-training MFU at Just ~12%
eliebakouch · x · 2026-10-06
Reflection released a new model under Apache 2.0 with a tech report. Analyst Elie Bakrouch estimates the pre-training ran at a surprisingly low 12% BF16 MFU, assuming 80% goodput over 4 weeks.
The architecture uses global/SWA interleaving at roughly a 3:1 ratio with slightly lower parameter sparsity (5%) than other OSS models. Training was notably stable, and the base model beats DSV4 on held-out code perplexity. He still calls it a win for the Western open-source ecosystem.
Related event: Reflection Open-Sources New Model Amid Training Efficiency Debate(2 posts)→
More from Infra
- ~90% of frontier lab compute now goes to post-training and inference — IanAndrewsDC · 2026-10-06
- Qwen 27B on 2× RX 7900 XT: 66.5 TPS single-stream, still short of claimed 100+ — EqualCryptographer67 · 2026-10-06
- Dev burns 842B tokens in September — $409k at API list price, pays just 3.4% via subscription — doodlestein · 2026-10-06
- One Dot burns 1.6B tokens/day on a $100 subscription — roughly $540k/month in API-equivalent compute — DarthSilent · 2026-10-06
- Charles Frye (Modal) explains inference engines: schedulers, KV cache, CUDA graphs, speculative decoding — AI Engineer · 2026-10-06
- Anthropic moves Claude Cowork fully to the cloud, dropping the battery-hungry local VM — Simon Willison · 2026-10-06