Originally published on rohitraj.tech
Week 32 of 2026 was a compression week. MiniMax-H3 shipped open weights that fall from 123.6 GB to 42.5 GB, Cloudflare doubled Kimi K2.6 context to 1.37M tokens with an FP8 KV cache, and AirLLM hit 27.5k stars streaming a 2.8T model through 4 GB of VRAM. Meanwhile JFrog found 54 of 55 SQLite CVEs from one repo were AI-generated fiction — one briefly scored a CVSS 10.0.
Read the full version with code samples, diagrams, and architecture details: This Week in AI Dev: Everything Got Smaller, and Six CVEs Turned Out to Be Fake (Week 32 of 2026)
More engineering notes: rohitraj.tech/en/notes
Top comments (0)