Overview
10 stories in this issue. The first 3 are today's priorities.
Hot model updates
- Top · Gemini 3.5 Live Translate covers 70+ languages
- Top · Memento 3 clears all 25 ARC-AGI-3 public games on frozen Claude Opus 5
Global AI news
- Top · QGIS enhancement proposal would ban AI-assisted contributions
- Tencent Yuanqi AI agent platform shuts down November 9
- Amazon reportedly in talks to buy Decart for up to $7B
- SWI Group raises €800M to expand AI compute in Europe and the US
- Karpathy: the gap in shared understanding of LLM capability is widening
Regional and early signals
- Zhipu's GLM-5.3-Highspeed in gray testing, reportedly ~490 tokens/s
- Orbbec exits lawn-mower robot ODM business
- Feimoyi raises nearly $100M seed for open-world "Physical Agents"
Jiufeng graphic based on the sources cited in this issue.
Hot model updates
01/10
Gemini 3.5 Live Translate covers 70+ languages
Google's Gemini 3.5 Live Translate supports more than 70 languages, keeps the speaker's intonation, and is available in the Gemini Live API preview.
On October 12, MarkTechPost compared 12 AI live translation tools against their official product, pricing and documentation pages. Google Translate's Live translate was one of them. According to the roundup:
- Model: Gemini 3.5 Live Translate
- Language coverage: 70+ languages
- Output: keeps the speaker's intonation
- Latency: Google says output runs a few seconds behind the speaker
- Where to find it: the Live translate feature in Google Translate, plus the Gemini Live API preview
For comparison, the same roundup lists Palabra.ai, which offers two-way speech-to-speech translation and claims under 1 second of latency. Its API costs $0.04 per minute, and Meetings plans start at $60 per month.
Limitations: MarkTechPost's piece is a roundup, not a Google launch announcement from that day. Pricing for Gemini 3.5 Live Translate has not been disclosed, and API access is still in preview.
Source: Google blog · MarkTechPost
02/10
Memento 3 clears all 25 ARC-AGI-3 public games on frozen Claude Opus 5
Huawei Noah's Ark Lab's agent scored 100 RHAE on the ARC-AGI-3 public set without training the underlying model.
Huawei Noah's Ark Lab built Memento 3 with UCL and other partners. The agent writes its own rulebook and a code world model. Its base model, Claude Opus 5, stayed frozen throughout, with no weight updates. It cleared all 25 games in the ARC-AGI-3 public set and scored 100 RHAE.
Limitations: The result is for the public set only. The report gives no results on non-public test sets. Our material comes from a single report and doesn't include the paper, the code or the full list of partners.
Source: Pandaily
Global AI news
03/10
QGIS enhancement proposal would ban AI-assisted contributions
A pull request to ban AI-assisted contributions has appeared in the proposals repository of QGIS, the open-source GIS project, and it drew a fast reaction on Hacker News.
The pull request is #387 in the QGIS-Enhancement-Proposals repository. On Hacker News it was posted as "QGIS bans AI assisted contributions" and reached 60 points and 46 comments within 37 minutes. One commenter said maintainers are already overloaded, and few contributors now save more time than a maintainer using an LLM and reviewing the output. Another suggested contributors should write detailed issues and let maintainers feed them to AI tools themselves.
Limitations: The available material doesn't show whether the proposal has been voted on or merged, or exactly what it covers. The Hacker News title is the submitter's wording; the PR text is the authority.
Image source: GitHub; mirrored on Jiufeng R2.
Source: GitHub PR #387 · Hacker News
04/10
Tencent Yuanqi AI agent platform shuts down November 9
Tencent is closing its Yuanqi agent platform, and agents published to WeChat accounts will stop replying.
The shutdown timeline:
| Date | Change |
|---|---|
| 2026-10-29 | No new agents or API keys can be created |
| 2026-11-09 | Platform goes dark; agents published to WeChat accounts stop replying |
| 2027-01-09 | Data export deadline |
Limitations: The report summary gives no reason for the shutdown and doesn't say whether Tencent will offer a migration path. This item has a single report, and our material doesn't include a link to Tencent's official notice.
Source: Pandaily
05/10
Amazon reportedly in talks to buy Decart for up to $7B
According to the WSJ, Amazon is in talks to acquire Decart, which makes software that lets AI models run on different processors.
The WSJ, citing people familiar with the talks, reported that a deal could close within days if talks hold and no other buyer appears. Decart raised $300 million at a $4 billion valuation in May, bringing its total funding to more than $400 million. Its investors include Radical Ventures, Nvidia Ventures, Adobe Ventures, Sequoia Capital and Benchmark. The company was founded in 2023 by brothers Dean and Orian Leitersdorf and Moshe Shalev.
Limitations: The deal is still in negotiation, and neither company has confirmed it. The report says talks could break down or another bidder could emerge.
Source: SiliconANGLE
06/10
SWI Group raises €800M to expand AI compute in Europe and the US
SWI Group is raising €800 million at a €6.1 billion valuation for data centers and AI compute in Europe and the United States.
The money comes from institutional investors through a new share issue by SWI Capital Holding, the group's listed holding company. SWI Group is led by founder and CEO Max-Herve George. The raise lifts its total equity raised since March above €1 billion and its free float to about 20%.
Limitations: RuntimeWire notes that the investors are unnamed and operating metrics are undisclosed, so the deal's terms and execution can't yet be judged.
Source: RuntimeWire
07/10
Karpathy: the gap in shared understanding of LLM capability is widening
Andrej Karpathy resurfaced one of his tweets from April, saying the gap in shared understanding of LLM capability is "widening," and gave rough head counts for different kinds of users.
Karpathy's rough estimates from the tweet:
| Group | People | Share |
|---|---|---|
| Barely used LLMs | ~6B | ~75% |
| Casual users | 1–2B | ~20% |
| Pro use in math/code | ~20M | 0.2% |
The ~20M group refers to people using frontier-grade LLMs professionally in math and code work. He adds that many professionals outside math and code fall into the ~20% casual group. The Hacker News thread reached 503 points and 732 comments within 15 hours. One commenter said many people in non-tech office jobs tried ChatGPT once or twice a year ago and still believe their work is too complicated to automate.
Limitations: The figures in the table are Karpathy's own rough estimates. Views in the Hacker News thread are individual commenters' opinions.
Source: Karpathy on X · Hacker News
Regional and early signals
08/10
Zhipu's GLM-5.3-Highspeed in gray testing, reportedly ~490 tokens/s
A new Zhipu model is rolling out at random inside Zcode, and a user reports output of about 490 tokens/s. At that speed, generating 10,000 tokens takes about 20 seconds (Chinese-language source).
According to user @leo114119, the model is available in Zcode 3.15.1 and assigned at random to subscribers. The model catalog seen by QbitAI lists:
- Context window: 1M tokens
- Max output: 131,072 tokens
- Thinking effort: low, high and max
Speeds of Zhipu's fast models as compiled by QbitAI (measured in different ways, for reference only):
| Model | Output speed (tokens/s) | Basis |
|---|---|---|
| GLM-5.3-Highspeed | ~490 | user report, gray test |
| GLM-5.1-highspeed | 400 | stated at May launch |
| GLM-5.3-FlashX | up to 200 | stated at launch last month |
| GLM-5.3-Flash | ~60 | Artificial Analysis, Z.ai API |
Limitations: The speed figure comes from a user leak, not from Zhipu, and nobody has verified that it holds up over time. Some Reddit users guess it is linked to Omen Alpha, an anonymous model on OpenCode, but that is speculation.
Source: QbitAI (Chinese-language source)
09/10
Orbbec exits lawn-mower robot ODM business
Leiphone reports exclusively that Orbbec decided internally in mid-September to stop building lawn-mower robots for other brands (ODM). Its vision module business is unaffected (Chinese-language source).
The ODM project won orders from a Shanghai customer but shipped only a few thousand units in total. Insiders give three reasons. First, on July 28 the US FCC added "advanced robotic equipment produced abroad" to its controlled list, and on September 11 it tightened its device authorization rules. Second, the market is crowded. Third, the customer's sales volume couldn't support the product line. Orbbec will now focus on data-collection equipment for embodied AI.
Limitations: This is a single exclusive report based on unnamed insiders, and Orbbec has not publicly confirmed it.
Source: Leiphone (Chinese-language source)
10/10
Feimoyi raises nearly $100M seed for open-world "Physical Agents"
Feimoyi, a robotics startup founded by former DeepRoute.ai executive Liu Nianqiu, raised a seed round led by IDG Capital and plans multi-site robot deployments next year (Chinese-language source).
Other investors include Yunqi Partners, Didi, Fosun RZ Capital, Yaotu Capital and Hesai. Liu says this year's work centers on collecting initial data. Deployments at multiple sites start next year to build a deploy–data–train loop, targeting commercial districts, office buildings and residential areas.
Limitations: The details come from an interview, and the date the round closed is not disclosed. The multi-site deployment is still only a plan for next year.
Write a prompt in your browser and get an image — 1K, 2K or 4K output, up to 15 reference images. 50 free generations on sign-up, no credit card.
Generate free
