In this issue
The short version
- Nvidia's updated DLSS 5 model runs up to twice as fast as the September version and uses about 20% less video memory at 4K, raising frame rates in NBA 2K27 by up to 1.4 times.
- MiniCPM-V 4.7's 35B-A3B version is an open mixture-of-experts model that ModelBest says runs on a phone with about 3 GB of memory at over 13 tokens a second.
- IDC says Nvidia's share of China's data-centre AI chip market fell to 34% in the first half of 2026, while Huawei rose to 28%.
- Character.AI filed for Chapter 11. Disney opened the auction with a $20 million bid for a licence to some of its IP. Export your chats if you use it.
October 14 covered AI running closer to the user, inside graphics cards and on phones, alongside a shift in China's chip market and the collapse of one of the best-known consumer AI start-ups. Each section explains what happened and what to do.
01
DLSS 5 gets a faster model
What Nvidia announced
Nvidia released a faster model for DLSS 5, its AI rendering technology that uses a neural network to generate and enhance game frames. The company says the updated model is up to twice as fast as the version launched in September, about ten times faster than the March demonstration, and uses about 20% less video memory at 4K. In NBA 2K27, the first game with DLSS 5, Nvidia reports frame rates up to 1.4 times higher. DLSS 5 runs on GeForce RTX 50 series cards. Nvidia also said DLSS 5 will come to RTX Remix, its tool for modders remastering classic games, on October 23, and that more games will follow.
What it means beyond gaming
DLSS shows a pattern that applies to AI generally: the same model becomes much cheaper to run within months as engineers optimise it. If a model feature seemed too slow or expensive when it launched, test it again a few months later. For gamers with RTX 50 cards, update drivers and the game, then compare frame rates with DLSS 5 on and off at your usual settings; image quality is a matter of taste, so judge it yourself rather than from screenshots.
Get HAI Daily in your inbox
The day's AI news, what it changes, and which tools to try.
02
MiniCPM-V 4.7: a large multimodal model on a phone
What ModelBest released
ModelBest's (面壁智能) OpenBMB team released MiniCPM-V 4.7, a multimodal model family for devices, in two sizes: 35B-A3B and 1B. The larger one is the series' first mixture-of-experts model, with 35 billion parameters of which about 3 billion are active at a time. ModelBest says it is the first open-source MoE model to run smoothly on a phone, thanks to a technique that predicts and loads the experts needed for the next step in advance, combined with quantisation. It reports over 13 tokens per second on an iPhone 17 Pro Max using about 3 GB of memory. The models are released under the Apache 2.0 licence, support iOS, Android and HarmonyOS, and come in several compressed formats. Benchmark comparisons are ModelBest's own.
When an on-device model makes sense
A model that runs on your phone works offline, keeps images and documents on the device and has no per-use cost. It suits tasks such as reading receipts, describing photos, translating signs or summarising a screenshot, where privacy or connectivity matters more than top quality. For complex reasoning or long documents, a cloud assistant such as ChatGPT, Claude or Qwen (通义千问) will still do better.
If you build apps, test an on-device model on your three most common image tasks and compare accuracy and speed with your current cloud API. Measure battery use and app size too, because a 3 GB model is a large download for users.
03
Nvidia's China AI chip share falls to 34%
What IDC found
According to IDC data reported by Reuters, Nvidia's share of China's data-centre AI chip market fell to 34% in the first half of 2026 as export controls limited supply. It shipped 678,000 AI accelerators in China, down 44% from a year earlier. Huawei (华为) shipped 572,000, raising its share to 28%. Alibaba's (阿里巴巴) T-Head (平头哥) held 7%, Cambricon (寒武纪) 5%, Baidu's (百度) Kunlunxin (昆仑芯) 4% and Hygon (海光信息) 3%, and nine of the 11 Chinese chipmakers IDC tracks increased shipments. IDC did not say which Nvidia products the figure includes.
What it means for users of Chinese AI services
More Chinese AI services now run on domestic chips. For users, the effects show up in which models are offered, how fast new models arrive and how prices move. If you depend on Chinese AI platforms, watch for providers announcing support for domestic hardware, and test speed and reliability after major platform changes. If you follow the chip industry, tools in the Finance AI category help track the companies involved.
04
Character.AI files for bankruptcy
What happened
Character Technologies, the company behind Character.AI, filed for Chapter 11 bankruptcy in Delaware on October 12. Founded in 2021, it was one of the most-watched AI chatbot start-ups of 2023. In 2024 Google paid to license its models, and its co-founders and a group of researchers joined Google. Since then the company has faced regulatory scrutiny and lawsuits over its effects on teenagers' mental health. Court papers reportedly show monthly losses of $8 million to $9 million.
Disney is the "stalking horse" bidder, offering $20 million for a non-exclusive, perpetual licence to selected IP and related claims, not for the whole company; other bidders can top the offer. The sale is expected to finish by late November or early December, and the company says the platform will be wound down if no buyer emerges.
What users and businesses should do
If you use Character.AI, export or copy any conversations and characters you want to keep now, and do not assume the service will continue. More broadly, the case is a reminder that AI products from start-ups can disappear. For anything you rely on, check whether you can export your data, and keep a copy outside the service. Parents should review which AI chat apps their children use and the age settings available. Our Character.AI guide explains its settings; Poe is one alternative that offers custom bots across several models.
What to try next
Put one change into practice
Retest any AI feature you rejected as too slow a few months ago. Try one on-device model for a privacy-sensitive task. Export your data from any AI service run by a start-up you depend on.
Explore Chatbots, Image Generation and Finance AI to build your shortlist.
How we put this issue together
The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.