Daily Briefing

China AI & robotics developments, rewritten weekly-context-free for US readers. Published daily at 07:00 Beijing time.

Sat, Oct 24, 2026

Moonshot AI releases Kimi K3.1 Pro, undercuts Opus 5.5 on price

Moonshot AI, the Beijing developer of the Kimi chatbot, released its K3.1 Pro model on Oct. 23 with a modest version bump but a larger price cut, positioning it against Anthropic's Opus 5.5 in the crowded frontier-model market.

The company announced the release in a 24-second video posted to X and its WeChat video channel on Oct. 23, according to the Chinese tech outlet Geek Park. The report frames the update as "more capability, lower price" — a challenge to Opus 5.5. Moonshot AI, founded in 2023 by former Tsinghua researcher Yang Zhilin, raised a reported roughly $1 billion round in 2024 and has leaned on its Kimi assistant to compete with DeepSeek, Alibaba and ByteDance at home. The input does not detail benchmark scores or specific pricing, so the scale of the claimed advantage over Opus 5.5 remains unverified.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
Moonshot AI launches Kimi K3.1 Pro with 3T parameters

Moonshot AI, the Beijing developer of the Kimi chatbot, released its Kimi K3.1 Pro model early on Oct. 24, a roughly 3-trillion-parameter system supporting a 1-million-token context window, according to the company. The launch intensifies China's frontier-model race, where context length and agentic benchmark scores have become key selling points for US-facing developers and investors.

Moonshot said the model scored 65.7 on the Terminal-Bench 4.0 evaluation, a test of command-line agent tasks. The company did not disclose training costs or parameter-count methodology. Moonshot AI, founded in 2023 by Yang Zhilin, raised funding from Alibaba and Tencent and has positioned Kimi against models from DeepSeek and Alibaba's Qwen. The article's title, roughly "The Tragedy of Kimi," suggests the piece questions the model's reception, but the provided excerpt does not detail those criticisms.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
Kimi and DeepSeek take opposite tacks on AI hiring

A tech-industry observer who says he has spoken with core teams at both Kimi developer Moonshot AI and DeepSeek described sharply contrasting hiring standards at the two Chinese AI startups, according to a post circulating on social media and reported by TMTPost. The comparison highlights how talent strategy, not just compute and data, is becoming a differentiator in China's crowded large-model race.

Moonshot AI, based in Beijing, develops the Kimi chatbot and has raised backing from Alibaba and Tencent; DeepSeek, also known as DeepSeek AI, is the Hangzhou lab spun out of quant fund High-Flyer. The post, whose author was not fully identified in the TMTPost summary, said one company prioritizes proven industry veterans while the other seeks young standout researchers. The report did not give specific hiring numbers or name individual recruits. TMTPost framed the contrast as a window into how China's model developers are positioning themselves as competition intensifies. Neither company has publicly commented on the claims.
Source: tmtpost.com · via Physix Frontier briefing desk · Editorial briefing card →
Moonshot AI releases Kimi K3.1 Pro, its largest model yet

Moonshot AI, the Beijing developer of the Kimi chatbot, released Kimi K3.1 Pro on Oct. 24, calling it the first model in the K3.1 series and its most capable to date at 3 trillion parameters, according to the company's official WeChat account.

The launch follows the earlier K3 flagship model. Moonshot, founded in 2023 by Yang Zhilin, has positioned Kimi as a long-context assistant competing with Alibaba's Qwen and DeepSeek. The report by Huxiu frames the release around a constraint rather than a capability: the company is pursuing many product directions at once while its access to compute remains limited. No specific figures on chip supply or spending were given in the excerpt.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →

Fri, Oct 23, 2026

BYD teases interior of second-gen Seal 07 sedan

BYD released official interior images of its second-generation Seal 07 on Oct. 23, showing a redesigned cabin with a large floating center screen and digital instrument cluster. The reveal previews how China's largest EV maker is refreshing its mid-range lineup amid intensifying price competition at home.

The new interior keeps the large floating touchscreen and digital gauge cluster but changes the steering wheel, center-console trim and air-vent shapes, according to IT Home. BYD said the seats support zero-gravity recline, massage, ventilation and heating, and that the rear row offers the largest space in its segment. The company also listed a 20-speaker Devialet audio system. BYD, based in Shenzhen, sells the Seal 07 as a plug-in hybrid sedan in China. No pricing or launch date was disclosed.
Source: ithome.com · via Physix Frontier briefing desk · Editorial briefing card →
Why founders invoke Anta and Li Ning when investors cash out

A Chinese sports-business commentary examines why Anta and Li Ning, the country's two largest domestic sportswear brands, surface so often in global deal rumors, and what that pattern signals when investors seek to exit.

The article, published by sports-business outlet Lanxiong Sports and reposted by Huxiu, notes that Anta and Li Ning have appeared as potential buyers in reported talks around Puma, Reebok, Under Armour's split with Stephen Curry, and Kylian Mbappe's Nike contract. Anta Sports, based in Jinjiang, Fujian, and Li Ning, founded by the Olympic gymnast of the same name, are China's leading homegrown sportswear groups. Their repeated mention in sale rumors is often attributed to their acquisition appetite and cash reserves. The piece frames the pattern as a signal for founders: when investors want to cash out, invoking these two names can raise the perceived value of an asset. No transaction is confirmed in the input.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
Great Wall's Wey brand sets Oct. 27 debut for second-gen Blue Mountain SUV

Great Wall Motor said its premium Wey marque will give the second-generation Blue Mountain SUV a global debut on Oct. 27, equipped with the Hi-4 plug-in hybrid system built around a 2.0T engine, according to the company.

The vehicle adopts a boxy, upright design, the company said in an announcement carried by Chinese tech outlet ITHome on Oct. 23. Great Wall Motor, based in Baoding, Hebei province, is one of China's largest SUV makers and has pushed plug-in hybrids as a bridge between gasoline models and full EVs. Wey is Great Wall's upscale brand, launched in 2016 and named after founder Jack Wey. The Blue Mountain is its flagship family SUV, competing in China's crowded large-SUV segment against models from Li Auto and Aito. Great Wall did not disclose pricing, range or market availability for the new model.
Source: ithome.com · via Physix Frontier briefing desk · Editorial briefing card →
Chinese court rules car's auto-call to police is not a confession

A Zhejiang driver charged with drunk driving argued that his vehicle's automatic emergency call to police amounted to turning himself in, but a court rejected the claim, according to the Chinese tech outlet ITHome. The case tests how Chinese law treats evidence and legal acts generated by connected-car systems rather than by their owners.

The driver had set "110" — China's police emergency number — as an emergency contact, ITHome reported on Oct. 23. When the car detected an anomaly, it reported to the automaker's emergency rescue center, which called police; officers then caught the man at the scene. He contested the charge, arguing the automatic alert should count as a voluntary surrender, a mitigating factor under Chinese criminal procedure. The court disagreed, holding that a system can notify authorities on a driver's behalf but cannot perform the legal act of self-surrender. The ruling adds to a small but growing set of Chinese cases weighing liability and intent when vehicle software acts autonomously. ITHome did not name the driver, the court or the automaker involved.
Source: ithome.com · via Physix Frontier briefing desk · Editorial briefing card →
Security experts warn against giving AI agents bank login access

A CNET report cautions users against handing banking credentials to autonomous AI agents, citing risks as tools like Meta's Muse, OpenClaw and OpenAI's dots gain traction. The warning lands as US consumers increasingly delegate web tasks to agents that can act on their behalf.

The article, surfaced on Hacker News, names Meta's Muse, OpenClaw and OpenAI's dots as examples of the agent trend. It does not detail specific incidents or vulnerabilities. Agentic tools that browse and transact require stored credentials to operate, expanding the attack surface if a session is hijacked or a model misbehaves. The piece offers no technical specifics or company responses. Readers should treat the guidance as general caution rather than a disclosed breach, according to the report.
Source: cnet.com · via Physix Frontier briefing desk · Editorial briefing card →
Kia targets Chinese EVs in Canada with sub-$40K EV3

Kia plans to position its EV3 compact electric SUV below the $40,000 mark in Canada, aiming to counter low-cost Chinese EV imports, according to InsideEVs. The move signals how global automakers are adjusting pricing and product strategy as Chinese brands push into new markets.

The EV3 is Kia's entry-level electric SUV, already launched in some markets as a mass-market model. InsideEVs reports the Canadian strategy targets affordability, where Chinese EVs have gained ground. Chinese automakers such as BYD and others have expanded exports, pressuring legacy brands on price. Canada has considered tariffs on Chinese EVs, adding uncertainty to the competitive landscape. The report did not specify launch timing or final pricing. Kia has not confirmed details beyond the sub-$40,000 positioning cited by InsideEVs.
Source: insideevs.com · via Physix Frontier briefing desk · Editorial briefing card →
H2O.ai forks jdupes, a duplicate-file utility

H2O.ai, the open-source machine learning platform company, has published a fork of jdupes, a command-line tool for finding and acting on duplicate files through deletion, hard linking, symlinking, or block-level deduplication, according to the project's GitHub repository.

The repository, listed on GitHub under the h2oai organization, carries the same description as the upstream jdupes project: identifying and taking actions upon duplicate files. The listing does not state why H2O.ai forked the tool or what changes it plans to make. jdupes is a widely used utility derived from the older fdupes program, commonly employed by system administrators and developers to reclaim storage by replacing redundant files with hard links or symlinks. H2O.ai is best known for its open-source machine learning platform and its Driverless AI product. The GitHub page provides no release notes, commit history summary, or stated motivation for the fork beyond the repository name and description.
Source: github.com · via Physix Frontier briefing desk · Editorial briefing card →
China's first fully AI-generated feature film set for release

A feature film titled Sanxingdui, billed as China's first fully AI-generated movie, is scheduled to reach Chinese cinemas on Friday, according to promotional material from Bona Film Group. The release would mark an early commercial test of generative video tools in theatrical distribution, a step US studios and AI vendors are also exploring.

Bona Film Group is a major Chinese studio and distributor listed on the Shenzhen exchange, known for backing mainstream commercial titles. The film takes its name from the Sanxingdui archaeological site in Sichuan, a Bronze Age culture whose bronze masks have become a popular cultural reference. The company has not disclosed which AI models or vendors produced the footage, and no budget or box office projections were included in the promotional material. Details on runtime, cast and whether any live-action elements were used were also not provided. AI-generated video has drawn investment from Chinese and US firms alike, though feature-length theatrical output remains rare and has faced scrutiny over consistency and rights.
Source: globaltimes.cn · via Physix Frontier briefing desk · Editorial briefing card →
Kimi K3.1 Pro narrows gap between open and closed AI models

Moonshot AI's Kimi K3.1 Pro, an open-source model, now trails the leading closed models by roughly one month, according to a comparison published on ModelCompare. The shrinking gap matters for US developers weighing open-weight alternatives against proprietary APIs.

The report, dated October 2026 and surfaced on Hacker News, tracks how far the best open model lags the best closed one. It frames that lag as having compressed to about a month. Moonshot AI, the Beijing developer behind the Kimi chatbot, has released successive open models. The item does not specify benchmark scores, parameter counts, or licensing terms. The comparison's methodology and the specific closed models used as the baseline were not detailed in the source material.
Source: modelcompare.ai · via Physix Frontier briefing desk · Editorial briefing card →
US government orders removal of DeepSeek products from federal systems

The US government has issued an order requiring the immediate removal of DeepSeek products and any third-party services built using the Chinese company's AI models, according to TechRadar, amid mounting federal scrutiny of Chinese large language models.

The directive covers both DeepSeek's own offerings and third-party tools developed with its technology, per the report. DeepSeek, based in Hangzhou, released its R1 reasoning model in January 2025, drawing attention for performance at low training cost. The order reflects broader US concern over data security and Chinese government access to AI systems. TechRadar characterized some of the reaction as "unwarranted fear," though it did not detail the order's scope, issuing agency, or compliance deadline. No specific agency, effective date, or enforcement mechanism was named in the source material.
Source: techradar.com · via Physix Frontier briefing desk · Editorial briefing card →
Unitree's 60% gross margin faces first robot price war

Unitree Technology reported 1.699 billion yuan (roughly $234 million) in 2025 revenue and a 60.44% gross margin, well above Chinese robotics peers, as analysts question whether that pricing power can survive an industry price war.

The Hangzhou-based maker of quadruped and humanoid robots posted 672 million yuan (roughly $93 million) in operating costs and 1.027 billion yuan (roughly $141 million) in gross profit for 2025, according to the AD Finance report republished by Huxiu. Peer margins were lower: UBTech at 37.67%, Dobot at 46.49%, DEEP Robotics at 52.83% and Leju Robotics at 40.78%. Unitree, founded in 2016 by Wang Xingxing, gained attention for low-cost quadrupeds and humanoids such as the G1 and H1. The report questions whether its premium can hold as rivals cut prices.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
Chinese startup claims edge chip beats Nvidia DGX Spark on Qwen

A little-known Chinese hardware startup says its 40-watt, 16-core Cortex-A55 system-on-chip outperforms Nvidia's DGX Spark and Apple's Mac Mini running Alibaba's Qwen models, claiming 256 TOPS can beat 1,000 TOPS in prefill workloads.

The company's argument rests on prefill performance, the prompt-processing phase of inference, rather than raw TOPS ratings that typically favor Nvidia hardware. It did not provide independent benchmarks, and the claim has not been verified by outside testers. The comparison targets Nvidia's DGX Spark, a compact desktop AI machine, and Apple's Mac Mini, both popular with developers running local models. Qwen is Alibaba's open-weight model family, widely used in China. The startup's identity, funding, and shipping timeline were not detailed in the report. US readers should treat the performance claim as unverified vendor marketing until third-party results appear.
Source: techradar.com · via Physix Frontier briefing desk · Editorial briefing card →
Moonshot AI launches Kimi K3.1 Pro model

Moonshot AI, the Beijing developer of the Kimi chatbot, has released a new model called Kimi K3.1 Pro, according to The Information, extending a rapid cadence of updates as Chinese labs push frontier-class systems.

The report offered no benchmark figures, pricing or availability details, and Moonshot did not immediately publish technical documentation. Moonshot AI was founded in 2023 by Yang Zhilin, a former Tsinghua University researcher, and raised backing from Alibaba and Tencent. Its Kimi assistant became known for long-context processing. The release lands amid intensifying competition with DeepSeek, Alibaba's Qwen team and Zhipu AI, and against US models from OpenAI, Anthropic and Google.
Source: theinformation.com · via Physix Frontier briefing desk · Editorial briefing card →
Moonshot AI ships first model update since breakout moment

Moonshot AI, the Beijing developer of the Kimi chatbot, released its first AI model upgrade since a breakthrough that drew comparisons to DeepSeek, according to Bloomberg. The update signals China's startup field remains competitive on model releases.

Bloomberg reported the launch on Oct. 23, describing it as Moonshot's first upgrade since what it called the company's own DeepSeek moment. The report did not specify the model's name, benchmark results, or pricing. Moonshot AI is a Beijing-based developer best known for Kimi, its long-context chatbot. DeepSeek, a Hangzhou lab, drew global attention in early 2025 with low-cost models that pressured US AI valuations. Details beyond the release itself were not immediately available in the source summary.
Source: bloomberg.com · via Physix Frontier briefing desk · Editorial briefing card →

Thu, Oct 22, 2026

VW-Xpeng co-developed EV sells just 267 units a month

The first vehicle jointly developed by Volkswagen and Chinese EV maker Xpeng has averaged only 267 monthly sales from April through August 2026, according to ThinkerCar data cited by Reuters, raising questions about whether the partnership's engineering work is translating into consumer demand.

Volkswagen and Xpeng, the Guangzhou-based automaker founded in 2014, announced their technology collaboration in 2023, with VW investing roughly $700 million for a 4.99% stake. The deal was framed as a template for how foreign legacy automakers could tap Chinese software and EV platforms. The slow start contrasts with Xpeng's own branded models, which have posted stronger volumes. ThinkerCar's figures, reported Oct. 20, cover April to August 2026. Neither company has publicly attributed the weak sales to a specific cause. The result underscores the difficulty of converting platform-sharing agreements into showroom traffic in China's crowded EV market.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
Xpeng services margin hits 75% as rivals become customers

Xpeng reported second-quarter revenue of 19.74 billion yuan (roughly $2.7 billion), with its services and other segment posting a 75.1% gross margin versus 12.1% for vehicle sales, according to the company's Aug. 24 results.

Xpeng Motors is a Guangzhou-based electric-vehicle maker founded in 2014 by He Xiaopeng, a former Alibaba executive. It competes with Nio and Li Auto in China's crowded EV market. The article, published by Huxiu, examines how Xpeng's services business — which includes selling technology to other automakers that are also competitors — carries far higher margins than its core vehicle manufacturing. The piece does not name specific customers or disclose contract terms. Xpeng has previously said it licenses its advanced driver-assistance software to Volkswagen Group.
Source: huxiu.com · via Physix Frontier briefing desk · Editorial briefing card →
← NewerOlder →