Grok 4.7 misses its Saturday: Musk wants more RL, not a calendar
A few more days of reinforcement learning, Elon Musk said Friday — response-length penalties, task management, and a model that still gives up too early.
Short1 minEasy
Priya Shah1 min read
- xai
- grok
- grok-4.7
- release

Photo: icons8 / UnsplashUnsplash License
Saturday was the date the timeline had circled. Friday night Elon Musk moved the circle. Grok 4.7 “needs a few more days to cook,” he wrote — not because the weights are missing, but because the reinforcement-learning pass is still teaching the model bad manners.
The list is specific. The team penalized response length too hard, so the model bails on long work. Task management is sloppy. It is not rigorous enough about checking its own output. Those are not “we need another pretrain” complaints. Those are “we overshot the reward model” complaints, which is the most 2026 way to miss a ship date.
Crypto Briefing, citing reporting around the delay, puts 4.7 near 2.1 trillion parameters against 4.6’s roughly 1.5 trillion — a bigger student that will cost more to serve and, xAI has implied, run a shade slower. Meta and OpenAI are both in a September posting cadence. A two-day slip is not a scandal. A model that “gives up on hard tasks” would have been a worse Saturday than an empty one.


