Grok 4.6 and 4.7: What Elon Musk Actually Announced (and What He Didn't)

Grok 4.6: What Is Confirmed

Parameter Count

Musk explicitly confirmed the 2 trillion parameter model, stating it is the "2T model" that is "better than our 1.5T in every way" . It is a direct successor to Grok 4.5, which is based on the 1.5T V9 foundation model .

Training Data: SpaceX Engineering

On July 21–22, 2026, Musk announced that SpaceX's proprietary "world-class engineering data" would be used for supplemental training of the "2T run" . This data includes internal archives from building orbital launch vehicles, Starlink hardware, and manufacturing processes. Critically, data restricted by the U.S. International Traffic in Arms Regulations (ITAR) — such as rocket engine propulsion details and guidance systems — is excluded. However, materials science, manufacturing processes, and Starlink hardware design data are included .

This integration gives Grok 4.6 access to specialized physical-reasoning data that standard web scraping cannot replicate . Multiple analysts have noted this could give Grok a significant edge in engineering, robotics, and scientific problem-solving .

The Competitive Context: Moonshot AI's Kimi K3

Grok 4.6 is being released into an increasingly competitive landscape. Just over a week before Musk's announcement, Beijing-based Moonshot AI released Kimi K3 on July 16, 2026 .

Kimi K3 Key Specs

• 2.8 trillion total parameters using a Mixture-of-Experts (MoE) architecture with 896 experts (approximately 16 active per token) .
• 1-million-token context window and native vision capabilities .
• Open-weight release, with full model weights promised by July 27, 2026 . Moonshot describes it as the first open "3T-class" AI system .

Benchmark Performance

Moonshot AI claims Kimi K3 beat Claude Opus 4.8 and GPT 5.5 on coding and general agentic benchmarks . CNBC reported this directly, stating Moonshot's claims were that the model "beat Claude Opus 4.8 and GPT 5.5 — models that sit just behind Anthropic and OpenAI's leading-edge systems" . Multiple other sources confirm these benchmark claims .

Grok 4.6 is explicitly positioned to compete against Kimi K3 .

Grok 4.7: What Is Confirmed vs. What Is Speculation

Confirmed

• Name and release window: Musk's July 24 post .
• SpaceXAI's broader accelerated roadmap .

Not Confirmed

• Parameter count: Musk did not disclose a parameter count or training details for Grok 4.7 .

What the Broader Pipeline Shows

As of May–June 2026, xAI's Colossus 2 supercluster was running seven concurrent training jobs, including one 6-trillion-parameter model and one 10-trillion-parameter model . These were described as variants of "Grok 5," not Grok 4.7 . Grok 5 has been reported as targeting 6 trillion parameters using an MoE architecture, originally slated for Q1–Q2 2026 but repeatedly delayed .

There is no independent source that explicitly labels any 6-trillion-parameter training run as tied to Grok 4.7. The 6T+ training runs on Colossus 2 appear to be Grok 5 or foundation-model work .

Key Caveats at a Glance

The bottom line is clear: Grok 4.6 (2T parameters, trained with SpaceX engineering data) is confirmed for an early August release, and Grok 4.7 is confirmed only by name and a late August release window. The 6-trillion-parameter Colossus 2 training runs are real, but the evidence ties them to Grok 5, not Grok 4.7. Musk has not disclosed what size 4.7 will be.