Elon Musk confirmed Grok 4.6 targets August 7 with 1.5T parameters and improved SFT/RL, followed weeks later by the 2.1T Grok 4.7. Here's the full xAI roadmap and why the split matters.
Sai Meghana G
Software Engineer
Elon Musk just laid out xAI's roadmap for the next month: Grok 4.6 around August 7 with 1.5 trillion parameters and improved fine-tuning, followed weeks later by the 2.1-trillion-parameter Grok 4.7. If you were expecting a straight leap to a 2T monster, that's not quite what's landing first.
Here's the actual roadmap, straight from Musk's own post on X.
Musk confirmed Grok 4.6 is targeting a release around August 7, 2026 — just a few weeks after Grok 4.5 shipped on July 8. That's an aggressive cadence on its own.
But the headline isn't the parameter count. Grok 4.6 comes in at 1.5 trillion parameters, which is actually the same size class as Grok 4.5 — not the scale-up a lot of people were anticipating. Musk described it as carrying "significantly improved SFT & RL" — meaning the gains are coming from better supervised fine-tuning and reinforcement learning, not from throwing more raw scale at the problem.
In plain terms: 4.6 is xAI polishing what they've already got, not making the base model bigger.
Then comes the jump. Grok 4.7 is slated to follow "a few weeks later" — no fixed date yet, likely late August or early September — and this is where the parameter count actually moves, up to 2.1 trillion parameters.
Musk's own words: it'll be "better than 4.6 in every way, except slightly slower to serve," though he noted it should have better token efficiency to help offset that.
So the shape of it is:
| Grok 4.6 | Grok 4.7 | |
|---|---|---|
| Parameters | 1.5T | 2.1T |
| Target date | ~Aug 7, 2026 | A few weeks after 4.6 |
| Focus | Improved SFT & RL | Bigger base model, better everything except speed |
| Trade-off | None specified | Slightly slower inference |
It's a reasonable strategy. Post-training improvements (SFT and RL) are cheaper and faster to iterate on than a full pretraining run, so shipping 4.6 first lets xAI get real-world usage data and feedback while the bigger 4.7 model finishes training in the background.
It also keeps xAI shipping something new roughly monthly, which matters in a race where Moonshot AI just open-sourced the 2.8-trillion-parameter Kimi K3, and competitors like OpenAI and Anthropic aren't slowing down either.
Both models are being trained on xAI's Colossus cluster in Memphis, which has reportedly scaled past 500,000 GPUs — so compute isn't the bottleneck here, sequencing is.
A couple of honest caveats:
Still, the roadmap itself is refreshingly clear: 4.6 is the polish, 4.7 is the jump. If both land on anything close to schedule, it's going to be a busy August for anyone tracking frontier model releases.
If you're using Grok Build for coding, the fact that xAI is shipping two model upgrades within a month says a lot about how seriously they're investing in the Grok ecosystem. Whether 4.6's fine-tuning gains translate to meaningful coding improvements remains to be seen — but the pace is hard to argue with.
New AI tools, automation workflows, and course drops — straight to your inbox. Join 2,400+ builders.
A leaked checkpoint reveals Cursor is testing Composer 3 internally under the codename Vega, with 4 reasoning tiers, 6 model variants, and performance reportedly between GPT-5.4 and GPT-5.5.
Anthropic launched Claude Opus 5 — near-Fable 5 performance at half the price, with a Fast Mode that runs 2.5x faster and improved coding benchmarks. Here's what changed, what it costs, and who should switch.
The White House publicly accused Chinese AI lab Moonshot AI of distilling Anthropic's Fable model to build Kimi K3, with Treasury threatening sanctions and Entity List designations. Here's what happened and what it means for the AI industry.
