Musk says: Grok 4.6 will launch in two weeks, 4.7 in four weeks—2 trillion parameters going head-to-head with Kimi K3

SpaceXAI CEO Elon Musk threw a new release schedule on X on 7/24, with Grok 4.6 launching in two weeks and Grok 4.7 taking over two more weeks later. Grok 4.6 is a 2 trillion-parameter model, which expands about one-third from Grok 4.5’s 1.5 trillion, aiming to catch up to and even surpass Moonshot AI’s (Kimi K3) roughly 2.8 trillion parameters on capability, while maintaining Grok 4.5’s speed and token efficiency.
(Background recap: Musk said Grok 4.5 launches tomorrow: benchmarked against Opus but faster and cheaper, directly facing GPT-5.6)
(Additional context: SpaceX has officially renamed itself to SpaceXAI, with Musk tying space and AI into one company)

Key summary

• Musk posted on 7/24 that Grok 4.6 will launch in two weeks and Grok 4.7 in four weeks
• Grok 4.6 is a 2 trillion-parameter model, about one-third larger than Grok 4.5’s 1.5 trillion
• The goal is to catch up to Moonshot AI’s Kimi K3 at roughly 2.8 trillion parameters, focusing on token efficiency

SpaceXAI CEO Musk on Friday (7/24) posted a new round of version timing on X. The post consists of just one line: Grok 4.6 in two weeks, and Grok 4.7 in four weeks. It’s less than three weeks until Grok 4.5 officially goes public.

Musk also mentioned the specs of 4.6 earlier. It’s a 2 trillion-parameter model, about one-third larger than the current Grok 4.5’s 1.5 trillion, with initial pre-training expected to finish this week. After that, there will be an additional layer of supplementary training—feeding in the engineering data accumulated by SpaceX—while parts subject to ITAR (the U.S. International Traffic in Arms Regulations) are excluded. The intent is to further raise engineering and reasoning capabilities.

2 trillion parameters versus 2.8 trillion Kimi K3

Musk spelled out the hypothetical opponent plainly: Grok 4.6 aims to match or exceed Moonshot AI’s (Kimi K3). The latter’s parameter scale is rumored to be about 2.8 trillion. With 800 billion fewer parameters, it still needs to win. The edge comes from xAI’s two-year focus on token efficiency—solving the same question with fewer tokens to produce the answer. Grok 4.5 currently outputs at about 80 tokens per second, and the official claims token efficiency for some tasks is twice that of models in the same tier.

“Win despite smaller parameter count but still win on the scorecard”—that line is xAI’s most repeated message over the past two years. The real test is whether speed holds up after scaling to 2 trillion.

The two-week compression is about post-training

The schedule is the real message of the post. Finishing pre-training is only half the story; there’s still post-training, reinforcement learning, safety evaluations, and red-team testing. In the industry, this whole sequence is typically estimated at four to six weeks. Musk gave Grok 4.6 two weeks, and Grok 4.7 another two.

As for Grok 4.7’s specs, he didn’t mention a single word this time. External speculation puts parameter counts between 4 trillion and 6 trillion, but xAI has not confirmed any numbers.

常見問題

When will Grok 4.6 be released?

On July 24, 2026, Musk said on X that Grok 4.6 will be released in two weeks, putting the timeframe in early August. Grok 4.7 will follow in four weeks, around late August. xAI has not yet announced an exact launch date.

What’s the difference between Grok 4.6 and Grok 4.5?

Grok 4.6 is a 2 trillion-parameter model, about one-third larger than Grok 4.5’s 1.5 trillion. It also additionally ingests SpaceX engineering data for supplementary training, aiming to raise engineering and reasoning performance while maintaining Grok 4.5’s output speed of about 80 tokens per second.