SemiAnalysis: GLM 5.3's biggest leap came from post-training alone, not pre-training
ben_burtenshaw · x · 2026-10-18
SemiAnalysis revealed that GLM 5.3 was built entirely by post-training the previous model — described as a wake-up call.
Key points:
- Post-training is far more GPU-efficient than pre-training from scratch; companies that can't afford training from scratch can post-train on proprietary data and beat some closed models.
- Comparing GLM 5.3 vs 5.2 on cyber capabilities shows a "night and day" difference, all from post-training — meaning there's still significant juice left to squeeze from post-training.
More from Models
- Nine models, five coding tasks on one 16GB RX 9060 XT: deepseek-v4.1-flash hits 5/5 — Ok_Substance2327 · 2026-10-18
- OpenAI's 10-Page PDF Details 6-Layer Always-On Agent Architecture With Zero of 16,600 Attacks Breached — blaizedsouza · 2026-10-18
- DeepSeek V4.1 Flash matches Opus 5.5 on 3D test at $0.15 vs $3.85 per run — SumitGup · 2026-10-18
- MiniCPM-V 4.7 hits HF trending: 35B-A3B MoE runs on phones in ~3GB — lmoroney · 2026-10-18
- Hot take: inference marketplaces will underdeliver as 8GiB local GPUs reach frontier-grade intelligence — Saul_Loveman · 2026-10-18
- AI Benchmark Circle Jokes: 'Is Life Worth Living Without METR Time Horizons?' — scaling01 · 2026-10-18