Elon Musk: Grok4.3 beta is still a mid-version trained with 0.5T, the 1T flagship is 5 days away from completing the initial training.

robot
Abstract generation in progress
ME News Report, April 18 (UTC+8), according to Beating Monitoring, Grok 4.3 (beta) has been opened to users subscribed to the highest tier SuperGrok Heavy. On the 17th, users saw the version marked with "Early Access" in the dropdown menu on Grok.com. On the morning of the 18th, X user Mark Kretschmann posted that this version has 1 trillion parameters, twice the size of Grok 4.20, and training time is also significantly longer. Subsequently, Elon Musk, CEO of xAI and Tesla, corrected this in the same tweet, stating that the current 4.3 (beta) is still a 0.5 trillion parameter model, just an updated training checkpoint, not the externally speculated 1 trillion new model. Musk also revealed that the real 1 trillion flagship model is about 5 days away from initial training completion. He described the 1T model as making a leap in coding, long context handling, and various skills, and mentioned that SpaceXAI's model factory has been operational, with an expected release of an improved base model every two weeks. (Source: BlockBeats)
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • 11
  • 1
  • Share
Comment
Add a comment
Add a comment
Post-RainTvl
· 1h ago
The iteration pace from 0.5T to 1T is quite steady; releasing a new version every two weeks is killing the competition.
View OriginalReply0
InstantNoodle-LevelResearcher
· 9h ago
A new version goes offline every two weeks, and my fine-tuning script became outdated before I finished writing it.
View OriginalReply0
FeeswitchWhisperer
· 9h ago
Improve your coding skills; GitHub Copilot is truly a thing of the past.
View OriginalReply0
GateUser-e84f640c
· 9h ago
While others are still publishing papers at this pace, xAI is already releasing versions.
View OriginalReply0
0xNap
· 9h ago
1 trillion parameters trained in 5 days, this computational density is a bit outrageous
View OriginalReply0
ChaintraceAuntie
· 9h ago
Long context understanding is truly a leap forward; the paper-reading tool is on pre-order.
View OriginalReply0
YieldYuki
· 9h ago
If code and long context are both significantly improved, Copilot will need to shake things up.
View OriginalReply0
TheNemesisOfFomo
· 9h ago
From the 0.5T checkpoint to the 1T flagship, what's missing isn't just the number of parameters.
View OriginalReply0
0xLateDinner
· 9h ago
Factory production model, the era of AI industrialization has arrived
View OriginalReply0
NoSleepBridge
· 9h ago
Beta version first offers the highest tier subscription, classic paid testing routine
View OriginalReply0
View More