Qwen3.8-Max weights ship with strings + Grok 4.6 arrives

Alibaba put the weights for Qwen3.8-Max on Hugging Face — a mixture-of-experts model with 2.4 trillion total parameters and about 95 billion active per request, natively handling 262,144 tokens of context. The license is not open source. Any company running a model-as-a-service or AI work assistant business whose revenue tops $50 million over twelve consecutive months must obtain a separate license from Qwen before using the model commercially, and products above 100 million monthly active users must display the model name prominently. Reuters had reported a revenue share; the published terms instead require a negotiated agreement.

A preprint posted August 10 finds that the encrypted reasoning blocks the major providers hand back to clients are interchangeable across sessions, users and models — so injecting a strong model's trace into a weaker, less-guarded one makes it print the hidden text in the clear. The authors decoded 315,320 publicly scraped blocks and recovered 367 pieces of personally identifiable information along with 182 credentials. They disclosed the flaw before publishing and propose cryptographic fixes for client-side reasoning.

xAI released Grok 4.6 on August 12, scoring 61 on the Artificial Analysis Intelligence Index — the same figure as OpenAI's GPT-5.6 Sol. The gains over Grok 4.5 land in agentic coding, at 69.9 percent versus 66.7 percent on CursorBench 3.2 and 65.9 percent versus 54 percent on DeepSWE 1.1. Pricing is $2 per million input tokens and $6 per million output, with a fast variant at twice that. It is live in Cursor, Grok Build, the API, OpenRouter, Vercel and Cloudflare.

Twitch will let Amazon train generative models on creators' stream recordings by default, with the opt-out sitting in channel settings rather than the creator dashboard. Chief product officer Mike Minton was unusually direct about why: "If this was opt-in, nobody would opt in. That's honestly the answer." Nearly 3,000 users joined an official Twitch stream to object. Asked whether videos had already been used for training, Minton said he did not know.

Google DeepMind moved a sign-language-to-text model called SL2T out of research and into two shipping products, Gboard and Live Transcribe, on the Pixel 11 at no additional cost. It was trained on more than 100,000 hours across over 50 sign languages and reaches a zero-shot score of 70 on the BLEURT translation metric on the FLEURS-ASL benchmark, well above any previously reported result. Pose tracking runs on the device, so only geometric coordinates leave the phone. American Sign Language to English ships first, guided by an advisory committee of Deaf organizations.

Bloomberg reports that Cognition, maker of the Devin coding agent, is already discussing a new round at a $40 billion valuation — up from the $26 billion pre-money it set in May. The target is tied to reaching a $1 billion annualized revenue run rate; three months ago that figure stood at $492 million, with enterprise usage of Devin growing about 50 percent month over month for half a year.

The Swedish vibe-coding startup closed a $400 million Series C, co-led by Menlo Ventures and the EQT-managed Scaleup Europe Fund, at a $13.3 billion valuation — double the $6.6 billion it set in December. Annualized revenue reached $500 million in June. Lovable says more than 60 million projects have been built since its November 2024 launch, that apps made on it draw over 900 million monthly visits, and that it plans to reach roughly 450 staff this year.

DeepSeek promoted V4 Pro to general availability as the 0813 build, ending a preview that had run since April. The mixture-of-experts model carries a one-million-token context window and lists at roughly 44 cents per million input tokens and 87 cents per million output — around a fifth of what Grok 4.6 charges on the input side, and a reminder that the price floor for capable inference keeps dropping out of China.