Community Discussion · Policy

Grok 4.6 is Coming: Analyzing the Compute Costs Behind It

Fang An Fan ZiAug 52026/08/04 332 views

I've been running a few models recently, and just these past couple of days I saw Musk teasing again at the SpaceX earnings call that Grok 4.6 will launch next week, with 1.5 trillion parameters, focusing on improvements in supervised fine-tuning (SFT) and reinforcement learning (RL). Honestly, my first thought wasn't about how strong the performance would be, but rather how expensive the compute would be.

Policy

2 replies

Ctrl + Enter to reply

Feng sirAug 5

In principle, combining SFT and RL does improve model performance on structured tasks, but the cost of full-parameter fine-tuning is truly prohibitive. I'm curious if you considered parameter-efficient fine-tuning methods like LoRA during testing, and how effective they are for a 1.5 trillion parameter model.

Long JiAug 5

I just used DeepSeek-V4-Flash a few days ago to help a friend tweak a financial model, and it's pretty much what you said. Compute costs are a total dealbreaker for SMEs. Full-parameter fine-tuning on 1.5 trillion parameters... who can afford that? 😂