Two weeks. That is what Elon Musk gave us for Grok 4.6, posted on X on July 24, 2026. Grok 4.7 follows exactly two weeks after that. Do the math and Grok 4.6 lands around mid-August, with 4.7 arriving in early September. A YouTube roundup caught the timeline first, and several outlets confirmed it soon after. The same day brought two more AI stories worth your time: a vision-capable version of GLM 5.2, and Microsoft’s free, open-source image generator, Mage Flow.
From Grok 4.5 to 4.7: A Six-Week Release Sprint
Grok 4.5 only went public between July 8 and 16, 2026. It runs on a 1.5-trillion-parameter V9 foundation and scored 29.0% on the SWE Marathon benchmark, ahead of Claude Opus 4.8’s 26.0%. Pricing sits at $2 per million input tokens and $6 per million output tokens, per Next Big Future. Musk’s team is already raising the stakes. Grok 4.6 is expected to double up to 2 trillion parameters, and Musk has called it “better than our 1.5T in every way.” The same report speculates Grok 4.7 could scale to 4 trillion, maybe even 6 trillion parameters.
Initial pre-training for Grok 4.6 was set to wrap the week of July 20 to 27, 2026. Supplemental training comes next, pulling in SpaceX’s large corpus of engineering data (minus anything ITAR-restricted) to sharpen reasoning and technical performance. The goal, according to Brian Wang at Next Big Future, is to match or beat Moonshot’s Kimi K3, reportedly around 2.8 trillion parameters, while keeping the speed and token efficiency close to Grok 4.5. That’s not nothing: Grok 4.5 already runs at roughly 80 tokens per second and burns about half the output tokens of comparable models on SWE-Bench Pro variants.
No pricing yet for the new models, but Grok 4.5 set the tone. Expect SpaceXAI to hold the line around that same $2 in, $6 out range, maybe with a faster variant at a premium. Cursor already runs Grok 4.5 and should support 4.6 at launch, drawing from the generous first-party token pool that keeps costs down for Teams and Enterprise users.
Why SpaceX Data Could Give Grok 4.6 an Edge
No other AI lab gets to train on SpaceX’s proprietary engineering corpus. Crypto Briefing flagged this as a real edge. The data covers design specs, simulation outputs and operational telemetry spanning rocket propulsion, materials science and orbital mechanics. That’s dense, reasoning-heavy material, a useful supplement to the usual diet of web text and code repos. If the training goes to plan, the payoff is a model that’s sharper at structured problem solving and technical accuracy, exactly where most large language models still stumble.
GLM 5.2 Gets Vision, Borrowing from Kimi K2.6
The base GLM 5.2 was text-only. A company called Bastion just gave it eyes, fitting a vision encoder pulled from Kimi K2.6, the open-source model from Moonshot AI. That encoder, called Moon read, was paired with a trained projector that turns images into tokens the language model can actually process. The result is NVAP 4: 381 billion parameters, well under the original GLM 5.2’s 750 billion, and light enough to run on a single 24GB GPU. It also supports Nvidia Blackwell GPUs out of the box, according to the original video report.
That’s a real step forward for open-weight vision models. A 381B model that fits on consumer hardware opens up local image reasoning, document analysis, visual Q&A, all without shipping your data to the cloud. No independent benchmarks yet. But building on a proven vision encoder from Kimi K2.6 instead of starting from scratch suggests Bastion cared more about reliability than reinventing the wheel.
Microsoft’s Mage Flow: A 4B Open-Source Image Generator
Microsoft quietly dropped Mage Flow, a free, fully open-source text-to-image model at 4 billion parameters. It comes in several flavours: a base version, a Quality variant, a Turbo variant, plus dedicated edit models. Some are fine-tuned with reinforcement learning. The Turbo models trade some polish for speed. You can try the whole thing with no sign-up on a public Hugging Face Space.
One test prompt asked for a detailed fantasy world map: 12 kingdoms, rivers, mountains, roads, cities, a scale bar, a compass, a decorative border, hundreds of readable place names. Mage Flow Quality nailed it, every element intact, text legible. Turbo was faster but the labels came out more garbled. Run the same prompt through Queen Image 3.0 and you get clean text too, but Mage Flow’s 4B size makes it far lighter to run locally or fine-tune. Code and weights are already downloadable. Microsoft says more documentation is coming in a separate release.
An August Full of Frontier Releases
Grok 4.6 and 4.7 land within a month of each other. GLM 5.2 Vision ships as a compact open-weight option. Microsoft drops a capable 4B image model. That’s a lot of practical, deployable AI packed into the end of summer. If you’re a developer or a team wanting to test local vision reasoning, play with open image generation, or plug the next Grok into your coding workflow, you’ll have plenty to work through before September hits.
Frequently Asked Questions
When will Grok 4.6 and 4.7 actually arrive?
Elon Musk posted on X (July 24, 2026) that Grok 4.6 is coming in two weeks and Grok 4.7 in four weeks. That points to mid-August for 4.6 and early September for 4.7, though these timelines can shift without notice.
What makes the GLM 5.2 vision model different from other open models?
It adds image understanding to a large text model by reusing the Moon read vision encoder from Kimi K2.6. The 381B parameter size lets it run on a single 24GB GPU, and it includes optimizations for Nvidia Blackwell hardware, making it practical for local inference.
Is Microsoft’s Mage Flow really free for commercial use?
Yes, Mage Flow is released as a fully open-source model. The weights, code, and a hosted demo are available on Hugging Face. Microsoft has not announced any usage restrictions, and typical permissive open-source licences would allow commercial use, but users should check the accompanying licence file for details.
