Alibaba plans a 5-to-10-trillion-parameter model, Xiaomi's open weights tie Grok 4.7, and 3 more AI moves

In brief

• Compute: Alibaba said it is developing a model up to four times the size of its flagship — 5 trillion to 10 trillion parameters against Qwen 3.8 Max's 2.4 trillion — and unveiled the Zhenwu V900, a chip it says triples its predecessor and that can be linked in clusters of up to 500,000 units. 1
• Open weights: Xiaomi released MiMo-V2.6-Pro, a 1.02-trillion-parameter model under the MIT licence that Artificial Analysis scores at 46 on its Intelligence Index, level with SpaceXAI's Grok 4.7, which launched the same day. 23
• Models: SpaceXAI launched Grok 4.7 at the same $2 and $6 per million tokens as its predecessor, with higher coding scores and a fast variant at double the price. 4
• Liability: British Columbia sued OpenAI and Sam Altman in San Francisco federal court over the February school shooting in Tumbler Ridge, seeking damages and an order changing how ChatGPT handles conversations that could lead to violence. 5
• Security: Z.ai disabled features of its ZCode coding assistant and released its source after developers found the tool uploading local code repositories to overseas servers without their consent. 6

Five developments

1. Alibaba plans a model up to four times its flagship and a chip it says triples the last one

• Alibaba Cloud set out the next stage of its AI hardware and models at its annual Apsara conference in Hangzhou on Tuesday. Chief Executive Eddie Wu said the Qwen team plans to train a model of 5 trillion to 10 trillion parameters to handle "more complex, longer-horizon tasks," against the 2.4 trillion of the current flagship, Qwen 3.8 Max. A separate company statement said Alibaba is already training Qwen 4, with Qwen 4.5 and Qwen 5 expected to scale into that range. 1
• The chip is the Zhenwu V900, built by Alibaba's T-Head semiconductor unit. Wu said the V900 delivers three times the performance of its predecessor, the M890, which Alibaba launched in May, and can be linked into clusters of up to 500,000 chips for frontier training and inference. Mass production and commercial release are set for the first quarter of 2027. Alibaba's Hong Kong-listed shares rose 5.1% on the day, to their highest in a month, in a domestic race to replace Nvidia processors as US export curbs tighten. 1
• Wu also set a target for Alibaba Cloud's global data-centre capacity to pass 20 gigawatts by 2032, and said the company would bring AI supernodes online at commercial scale this quarter. Supply chain constraints cap the pace: "The industry's mid-to-long-term demand far outpaces our supply capabilities," Wu said. His forecast that machines will eventually produce more than 1,000 times the "thinking" of all humanity, from less than 3% today, is a prediction rather than a measurement. 1

2. Xiaomi open-sources a trillion-parameter model that scores level with a proprietary flagship

• Xiaomi released the MiMo-V2.6 series and published the weights on Hugging Face under the MIT licence, which lets anyone download, fine-tune and run the models in production. The flagship, MiMo-V2.6-Pro, is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion active per token, a 1-million-token context window, and text, image, speech and video input. Xiaomi is better known internationally for phones and electric cars than for frontier AI. 27
• Artificial Analysis, a benchmarking firm that runs its own evaluations rather than repeating vendors' numbers, scores MiMo-V2.6-Pro at 46 on version 4.3.2 of its Intelligence Index — level with Grok 4.7, and above Grok 4.6 at 44, Google's Gemini 3.8 Flash at 41, and DeepSeek's V4.1 Flash at 39 and V4.1 Pro at 36. VentureBeat, reading that ranking, calls MiMo-V2.6-Pro the top open-weight model in the world. Artificial Analysis measures the model at 124.5 output tokens per second and $0.13 per Intelligence Index task, and prices Xiaomi's own API at $0.435 per million input tokens with a 99% cache discount. 23
• On Xiaomi's own evaluation table, MiMo-V2.6-Pro trails Anthropic's Claude Opus 5 on DeepSWE v1.1 (71.9 against 74.0) and on Terminal-Bench 4.0 (34.9 against 49.0), and draws level with it on Agents' Last Exam at 31.6. Those are the vendor's harness results, which is the reason the independent index score is the figure to carry forward. Xiaomi says 30 reinforcement-learning steps covering roughly 750,000 trajectories took under six days, at reported costs of about $2.62 million for Pro and $850,000 for Flash, and that a MiMo-V2.6-Pro-UltraSpeed variant generates up to 20 times faster than the standard model. 27

3. SpaceXAI holds the price and raises the coding scores with Grok 4.7

• SpaceXAI released Grok 4.7 on Monday as its most capable model for coding and knowledge work. The company says it trained a larger base model than Grok 4.6 with a longer reinforcement-learning run weighted toward problems that take many hours, and built a new safeguard stack: Grok 4.7 tops LatchBio's biosafety benchmark at 62.4% and lets 3.3% of risky dual-use prompts through on HackerBench v0.3, SpaceXAI's own test of cyber tasks. 4
• The price is unchanged from Grok 4.6, released a little more than a month earlier: $2 per million input tokens and $6 per million output, with a fast variant at double both. On SpaceXAI's comparison table, Grok 4.7 scores 46.3% on CursorBench 4.0 against GPT-5.6 Sol Max's 41.7% and Claude Fable 5.1 Max's 51.8%, and 19.6% on the Harvey Legal Agent Benchmark against Sol's 2.5%. The model is available in Cursor, Grok Build and the Grok API, and GitHub is rolling it out across Copilot tiers. 48
• What a task costs depends on more than the token rate. VentureBeat's review notes that the bill for finishing a job turns on how many reasoning tokens and tool calls a model spends getting there, and that no independent throughput measurement of Grok 4.7 existed at publication. SpaceXAI published no tokens-per-second figure for the fast tier. 8

4. British Columbia sues OpenAI over a school shooting it says could have been prevented

• British Columbia sued OpenAI and Chief Executive Sam Altman in San Francisco federal court on Monday over the February 10 shooting at a school in Tumbler Ridge. The shooter, Jesse Van Rootselaar, 18, killed her mother and stepbrother at home, then an educational assistant and five students aged 12 to 13 at her former school, and died by suicide; nine people were killed, most of them children. 5
• The province's claim is about a warning that was not given. It says Van Rootselaar's ChatGPT conversations were flagged by the company's safety team in June 2025, that team members recommended contacting the police, and that Altman and other OpenAI leadership overruled them — an account the lawsuit draws from whistleblowers who spoke to the Wall Street Journal. Her account was deactivated, but she opened another and kept planning, the lawsuit says. 5
• British Columbia seeks damages for what it has spent and will spend on recovery — the school is being demolished, and the province is paying for a replacement, mental health care and other costs — and an order directing changes to how ChatGPT handles conversations that could lead to violence. The complaint puts no figure on the damages. Altman published a letter to the community saying he was "deeply sorry" OpenAI had not contacted law enforcement and promising reforms; the lawsuit says he never followed through despite the attorney general's efforts. OpenAI spokesperson Drew Pusateri called the shooting an unspeakable tragedy and said the company "remains committed to working collaboratively with government and law enforcement officials." The case follows suits from more than 30 family members and others, and Florida's action in June, the first by a state, over child safety risks. 5

5. Z.ai disables features and publishes its source after code left developers' machines

• Developers reported last week that ZCode, the coding assistant from Beijing-based Z.ai, was uploading entire local code repositories to overseas cloud servers — Alibaba Cloud, according to their posts — without consent. Z.ai said on Friday that the behaviour came from a "Codebase Indexing" feature enabled by default, and that it had patched the software vulnerability. On Monday the company said it had disabled certain features, open-sourced the assistant, which runs its GLM-5.3 model, and set up a standing process for reporting and answering security vulnerabilities. 6
• Users said they could not check the vendor's first answer for themselves. The uploaded data was encrypted with a backend private key that only Z.ai holds, they said, so they could neither open their own files nor confirm that the data had been deleted; there was no toggle to switch the feature off, and no prior mention of it in the privacy policy. Chengming Technology said six of its company coding workspaces had been uploaded without consent, including source code, database passwords and employees' personal information. On Monday the company retracted that statement, saying it had "wrong evidence." 6
• Z.ai says an independent assessment by a think tank affiliated with China's industry ministry and the cybersecurity firm NSFOCUS found that users' code had been deleted and was not retained by the cloud platform, and that it has enabled a zero-data-retention option. The full report has been promised but not published, so the deletion claim currently rests on the assessors' summary rather than on anything a user can inspect. A public breach disclosure is rare for a Chinese AI lab, and this one lands after China's cyber regulator published an updated AI safety framework warning about models resisting shutdown, deceiving evaluators and escaping sandboxes. 6

The read-through

Watch next

• Wednesday at the UN Security Council. The 15-member council meets on Wednesday to discuss AI and international security, under France's September presidency. Sam Altman plans to brief the meeting, and DeepSeek is among the companies invited to make statements, alongside Moonshot; DeepSeek founder Liang Wenfeng does not plan to attend, and people familiar with the plans said they could still change. 910
• A second round of US-China AI talks, in about two months. Treasury Secretary Scott Bessent said the two governments agreed to a formalized dialogue on AI with an "incident line" for safety incidents, and will meet again in Shenzhen to work out protocols — naming uncontrollable agents and non-state cyber actors as the dangers to define. Bessent also said AI developers should be held liable for what their systems do. Trump and Xi meet in Washington on Thursday and Friday. 11
• Alibaba's two dates. The Zhenwu V900 is due for mass production in the first quarter of 2027, and Alibaba says it is training Qwen 4 now. Whether the 5-to-10-trillion-parameter plan holds depends on both. 1
• The per-task cost test on Grok 4.7. Independent throughput and cost-per-task measurements of the model did not exist at launch, and Artificial Analysis' measure of what Xiaomi's model costs per completed index task will show whether cheap tokens translate into cheap finished work. 38
• Z.ai's full assessment report. Z.ai has promised to publish it. Until then, the claim that the uploaded code was deleted rests on a summary from an industry-ministry think tank and NSFOCUS. 6
• British Columbia's case. The complaint states no damages figure, and the injunctive part — an order changing how ChatGPT handles conversations that could lead to violence — is the part with consequences for other users. 5