DeepSeek V4 Could Continue Pretraining with MOPD Reusing Domain Experts
teortaxesTex · x · 2026-07-31
A discussion suggests DeepSeek V4 could continue pretraining from intermediate checkpoints, not just post-training; MOPD allows cheap reuse of domain experts. Also, a user noticed V4 Flash API fingerprint update, with knowledge cutoff updated to Nov 2025.
More from Models
- Anthropic Models' 'Tortured' Behavior Potentially Linked to Safety Training — repligate · 2026-07-31
- Leaked DeepSeek-V4 CoT Raises Eyebrows: Looks Like Direct Mandarin Translation — teortaxesTex · 2026-07-31
- Luna Offers Cheaper Inputs, but DeepSeek Wins Cache Economics in Long Agentic Sessions — teortaxesTex · 2026-07-31
- User Criticizes Hive AI: Fake Image Missed, Fake Sound Detected in Silent Clip — henkvaness · 2026-07-31
- DeepSWE Coding Agent Leaderboard: Claude Opus and GPT-5.6 Top the Charts — tristanbob · 2026-07-31
- Hive AI Detection Fails: Fake Crows Missed, 36% Music Detected in Silent Clip — henkvaness · 2026-07-31