THE 60-SECOND VERSION
● Qwen 3.8-Max weights shipped August 12, but under a custom license, not Apache. Text-only. Vision and the 1M context are API-only.
● Grok 4.7 finished initial pre-training and slipped to roughly 3 to 4 weeks out. xAI is now feeding it SpaceX engineering data.
● OpenAI files its preliminary injunction response in the Apple trade secrets case today. Hearing October 1.
● 14 days left on Claude Sonnet 5 at 2 dollars per million input. September 1 it becomes 3 dollars, plus a tokenizer change.
Qwen 3.8-Max: the first Max-class open weights, and the fine print
Alibaba published Qwen3.8-Max open weights on August 12 at the Hugging Face repo Qwen/Qwen3.8-2.4T-A95B, in BF16 safetensors plus an FP8 variant. NVIDIA's own deployment engineering blog confirmed the release the same day while walking through how to serve the model on a GB300 NVL72 rack.
This is the first time a Max-class Qwen model has been downloadable. It is also not the model the API sells.
WHAT GOT STRIPPED
The downloadable checkpoint is text-only with forced thinking. The hosted qwen3.8-max API keeps the multimodal input and the 1M-token context that made the product notable at launch. Same name, different artifact.
The license is the bigger story. It is a custom Qwen3.8-Max License, not Apache 2.0 and not MIT. The terms that matter:
- Keep the copyright and permission notice in all copies
- If your product exceeds 100M monthly active users or 20M dollars monthly revenue, you must prominently display the model name in your UI
- If you run a model-as-a-service or AI work assistant business with over 50M dollars trailing-twelve-month aggregate revenue including affiliates, you need a separate paid license before commercial use
- Purely internal use is carved out, provided you do not expose the model, its outputs, or its capabilities to third parties
For most teams reading this, none of those thresholds bite. For anyone building a product on top of it and selling access, they do. That is the distinction Apache 2.0 does not make and this license does.
Compare it to the sibling release: Qwen3.8-27B landed August 14 under plain Apache 2.0, no thresholds, no revenue riders, and it scored 61.7 on SWE-Bench Pro. The small model got the clean license. The flagship got the revenue-share.
We broke the full license terms and the weights-versus-API gap down separately.
Grok 4.7: pre-training done, and it slipped anyway
As of August 13, the xAI model list at docs.x.ai still topped out at grok-4.6. No grok-4.7 model ID, no pricing, no context window, no benchmark card.
What did change is that Musk confirmed on X that initial training is complete and the company has moved into a supplemental training run, feeding what he described as a large volume of SpaceX engineering data into the model, excluding ITAR-restricted material. The release window moved out to roughly three to four weeks, which puts it in early to mid September rather than late August.
This is an unusual reason for a delay. Most frontier slips are compute or safety evaluation. This one is a deliberate decision to post-train on a proprietary engineering corpus that no other lab has access to, and it tells you what xAI thinks the differentiator is: not general reasoning, but engineering and physical-systems work.
Meanwhile the model you can actually call is Grok 4.6, shipped August 12:
| Spec |
Grok 4.6 (shipped) |
Grok 4.7 (unshipped) |
| Parameters |
1.5T |
~2.1T reported, unconfirmed |
| Price per 1M |
2 in / 6 out, doubling above 200K input |
Not published |
| Context |
500K |
Not published |
| AA Intelligence Index |
61 |
No independent score exists |
| Status |
Live on API, Grok Build, Cursor, X, OpenRouter |
Supplemental training |
Every Grok 4.7 detail currently in circulation comes from Musk posts on X, not from an xAI model card or blog. Treat all of it as a founder timeline until docs.x.ai says otherwise.
OpenAI: injunction response due today, S-1 window still open
OpenAI's response to Apple's preliminary injunction motion in the trade secrets case is due today, August 17. OpenAI has already filed a 31-page motion to dismiss the underlying suit, arguing the claim is meritless and that Apple is litigating to compensate for its own difficulty shipping AI features. The hearing is set for October 1.
Separately, the public S-1 has still not appeared on SEC EDGAR. The confidential draft went in June 8. SEC review of confidential filings typically runs 60 to 90 days across multiple comment rounds, and the registration statement must be public at least 15 days before a roadshow starts. That math keeps the public prospectus in the late-August window and a listing in mid-September at the earliest, with Q4 2026 or even 2027 still live options.
The numbers everyone is waiting to see audited: roughly 2 billion dollars per month in revenue, reported losses of about 1.22 dollars for every dollar earned, an 852 billion dollar valuation from the March round, and the Microsoft revenue-share terms that have never been disclosed. Goldman Sachs and Morgan Stanley are leading. Not investment advice.
Deadlines still running
| Date |
What happens |
| Aug 17 (today) |
OpenAI preliminary injunction response due, Apple trade secrets suit |
| Aug 31 |
Claude Sonnet 5 moves 2 to 3 dollars per million input, plus a tokenizer change adding 10 to 35 percent tokens on code |
| Aug 31 |
kimi-k2.5 and moonshot-v1 sunset, migrate to kimi-k3 |
| Oct 1 |
OpenAI vs Apple hearing |
| Oct 24 |
deepseek-chat and deepseek-reasoner deprecated |
What we are watching next
The Qwen license is the thread to pull. If a Max-class frontier model can be downloaded but the license carves out anyone who resells access above 50 million dollars, that is a new middle category between open and closed, and it is aimed squarely at the model-router and AI-assistant businesses that would otherwise have been the biggest beneficiaries of open weights. Watch whether Meta, DeepSeek or Moonshot copy the structure.
Second: whether Grok 4.7 ships with any published benchmark at all, or whether it lands the way 4.6 did, announced on X first with the model card following later.
FAQ
Can I use Qwen 3.8-Max weights commercially?
Yes for most cases, with conditions. Above 100M monthly active users or 20M dollars monthly revenue you must display the model name in your UI. If you sell model access or run an AI work assistant business above 50M dollars trailing-twelve-month revenue including affiliates, you need a separate paid license first. Internal-only use is carved out.
Do the Qwen 3.8-Max weights include vision?
No. The downloadable checkpoint is text-only with forced thinking. Multimodal input and the 1M context window remain exclusive to the hosted qwen3.8-max API at 2 dollars in and 6 dollars out per million tokens.
When is Grok 4.7 actually launching?
No date has been published by xAI. As of August 13, Musk indicated roughly three to four weeks out with supplemental training underway, which points to early or mid September. There is no model ID at docs.x.ai yet.
Should I wait for Grok 4.7 before committing to a model?
Grok 4.6 is shipped, scored at 61 on the Artificial Analysis Intelligence Index, and priced at 2 dollars in and 6 dollars out below 200K input tokens. Grok 4.7 has no price, no context window and no independent benchmark. Build on what exists.
Has OpenAI filed its public S-1 yet?
Not as of today. The confidential draft went to the SEC on June 8, 2026. The public prospectus must appear at least 15 days before a roadshow, which keeps late August in play for the filing and mid September at the earliest for a listing.