Kimi
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 25
- 17-year-old Kimi researcher behind Attention Residuals leaves for Silicon Valley
A report republished by 36Kr says 17-year-old Kimi machine-learning researcher Chen Guangyu announced his departure in an Oct. 15 post on X and replied to Anthropic scalable reinforcement learning lead Sholto Douglas that he is heading to the Bay Area. He joined Moonshot AI as a 16-year-old intern and is one of the co-first authors of the Attention Residuals paper, whose Block AttnRes variant, proposed with Zhang Yu, splits networks into blocks and cuts memory complexity from layer-scale to block-scale; the article says it matched a baseline using about 80% of the compute on a Kimi Linear 48B model trained on 1.4 trillion tokens, with under 2% added inference latency. The piece says the technique is one of the foundations of Kimi K3, described as a 2.8-trillion-parameter, 3T-class open-weight model released in July 2026, and uses the move to examine whether Chinese AI firms can retain elite talent, drawing parallels to hiring battles in smart-driving and robotics.
36氪 人工智能 · 🔥 7 - Meta Muse sparks personal-agent race; Kimi pulls back as giants weigh payoffs
A 36Kr analysis says Meta Muse's personal-agent push has set off a wave in China: Alibaba is testing "Qwen Go!", Tencent is pushing LightVela and trialing Handy Bot, and ByteDance is adding automation to Doubao. By contrast, Moonshot's Kimi Claw, launched in February, is still in beta seven months on, with the author saying internal maintenance of the product has visibly slowed. The author attributes this to unit economics: a 199-yuan monthly consumer plan spreads to roughly 0.5 yuan per simple chat, but a complex agent task triggers many model calls, sandbox compute and long context, and per public information more than 70% of Kimi's revenue comes from token-billed business API usage. The piece argues the giants are instead banking on indirect returns — Meta on strong-intent ad targeting, Alibaba on GMV and take rates, Tencent on private-domain retention and payment flows, ByteDance on protecting its content and traffic entry — and that payment authorization is the endgame, noting Doubao's phone assistant was blocked by major apps when it tried GUI-based cross-app ordering and payment.
36氪 人工智能 · 🔥 6 - Open-weight agents 88% more susceptible to indirect prompt injection, study finds
Security research firm PromptArmor published a 9,800-trial evaluation claiming that agents built on open-weight models (DeepSeek, Kimi, GLM, Qwen) fell for real-world indirect prompt-injection attack chains 26.7% of the time, versus 14.2% for proprietary models (GPT, Claude, Gemini) — about 88% higher. It cites per-attack gaps such as 49% versus 9% for a chain that summarizes a URL and runs malware, and 36% versus 5% for a phishing login overlay. The study also says proprietary vendors' cost-effective models were 169% more susceptible than their flagships, while Qwen's and DeepSeek's cheaper models were slightly more robust than older flagships; PromptArmor has not released its full methodology or dataset, so the results are unverified.
PromptArmor Research · 🔥 7 - ModelSphere open-source platform targets production-grade AI deployment
On October 16, Paradigm Intelligence (范式智能) and Beijing Ruisizhilian Technology (睿思智联) announced ModelSphere (模域), a one-stop production-grade AI deployment platform that is free and open source for enterprise users under Apache 2.0 and is already available on GitHub. The platform combines compute management, model deployment and runtime optimization in one workflow, offering a graphical console, preset deployments, automatic elastic scaling, prefill-decode disaggregation, three-tier KV caching, high-availability load balancing and automatic service-resumption after stream interruptions, plus AutoTune scheduling intended to speed up first-token response. The vendors claim about 285% higher peak throughput and roughly 4x performance versus a baseline, figures that are company-stated and not independently verified; the platform bundles the EngineX inference engine to schedule heterogeneous Chinese accelerators from Huawei Ascend, Iluvatar CoreX, Moore Threads and Hygon DCU, and says it has optimized support for models such as DeepSeek-V4/V4-Flash, Qwen, Kimi, MiniMax and GLM.
InfoQ 中文 AI&大模型 · 🔥 7 - Chinese LLM makers push into AI hardware to grab entry points and revenue
On Oct 13, StepFun chairman Yin Qi unveiled the STEPX Neo agent phone in Shanghai, while Alibaba's Qwen released two AI glasses and an AI headset in Hangzhou the same day, with both events showcasing AI capabilities. Earlier, ByteDance and ZTE's Nubia NaviX Ultra — better known as the "Doubao phone" — went on sale Sept 16 with over 360,000 reservations on JD.com, and Tencent's WorkBuddy open platform launched Sept 2 in Shenzhen with nine co-branded hardware devices spanning glasses, earbuds, recording cards and speakers. A former executive at a leading Chinese LLM company is quoted saying the push is mainly about capturing entry points, since users show little loyalty to models, and that selling devices is another monetization path; listed Zhipu spent 2.131 billion yuan on R&D in the first half, about 2.2x revenue, while MiniMax spent US$297 million, roughly 2.5x revenue, with both loss-making. The report notes no LLM company has disclosed accurate AI hardware sales figures, and hands-on tests found weak real-time translation accuracy and an AI phone that failed to complete a takeout order in several minutes.
36氪 人工智能 · 🔥 6 - Doubao Phone Assistant opens screen automation; 155 apps opt out
Doubao Phone Assistant said the 30-day public notice period for its Screen Automation Execution Protocol (SAEP) has ended, so it will gradually open screen-automation access to third-party apps: those that pass risk assessment and have not explicitly objected can be controlled via simulated GUI taps and screen recognition. The refusal list published the same day grew from an initial 151 to 155 apps, including 66 from Tencent (WeChat, QQ, Honor of Kings, Tencent Yuanbao), 23 from Alibaba (Taobao, Alipay, Amap, DingTalk, Qwen), plus 14 from Kuaishou, 11 from Meituan, and AI peers DeepSeek and Kimi. Doubao said the refusals cover only agent-driven operations, not manual use of those apps on devices such as the Nubia M153 and NaviX Ultra, and that developers can opt out at any time, with support shut off within three working days.
IT之家 AI · 🔥 8 - First hands-on comparison of Meta's Muse Wonder
Arena shared a video in which reviewer @petergostev puts Meta's Muse Wonder (Max) through one-shot 3D scenes, animated HTML and front-end design tasks. He compares it against Astra, Opus 5.5, GPT-6.1 Sol, Claude Sonnet 5.5, Fable, Kimi and Mistral, and includes the cost of every generation. Arena says its scores for Muse Wonder are coming very soon, so this is currently only a first-impressions look.
arena (X) · 🔥 0 - MiniMax opens first Tmall flagship store selling M Plan subscriptions
Chinese model maker MiniMax has opened its first official brand flagship store on Tmall, listing its M Plan subscription in three tiers — Go, Explore and Build — available monthly or annually, according to multiple Chinese outlets reporting on Oct 13. AIbase said an opening promotion puts the Go monthly plan at about 24.5 yuan for a first order, a price cited by only one source. The reports also note that Tmall launched a token top-up hub called "AI Space Station" bringing together subscriptions from Alibaba Cloud, Zhipu, Kimi, MiniMax and DeepSeek, and that Zhipu and Qwen Office have already opened stores while Kimi and StepFun are in talks to follow. Model vendors had previously sold subscriptions mainly through their own websites.
IT之家 AI · 🔥 0 - Opinion: Just say no to OpenAI and Anthropic, use open weights
A personal blog post, upvoted past 100 points on Hacker News, argues that developers should stop relying on OpenAI and Anthropic and instead use open-weight models with private inference services. The author, a software and security engineer, writes that his vibe-based impression is that GLM-5.3 is roughly on par with Claude Opus 5.5 for software engineering, and describes using GLM-5.3, DeepSeek, Qwen, Kimi and Mistral Large 4 through Tinfoil, Maple AI and OpenCode Go. He criticizes the two US labs' closed models, strict cybersecurity guardrails and gated "trusted access" approval programs, and cites a Reuters report that Z.ai's ZCode coding harness silently uploaded local source code.
Hacker News · AI（100+ 分） · 🔥 0 - Jensen Huang: Chinese open models like Kimi are 'good enough,' backs open source
Speaking at a forum hosted by DBS Bank, NVIDIA CEO Jensen Huang said Chinese AI models including DeepSeek, Kimi and GLM are "already good enough." He described himself as a firm supporter of open source, calling it the best way to drive efficiency gains, and said the market will decide which models best fit different use cases. Huang also rejected the framing of AI competition as "a race with a finish line," saying both the US and China will be long-term players, and argued that cheaper, more open models will let more people and companies benefit from AI, increasing demand for NVIDIA chips, data centers and compute. The report says he argued the market first misread DeepSeek's impact and that Wall Street is now misreading Kimi's, and that quantum computing could reach mainstream adoption within three to five years.
cnBeta · 🔥 0 - Flo Health fine-tunes open models with Databricks to boost medical safety and usefulness
A Databricks blog post describes how Flo Health, a women's health app with 81 million monthly active users, built safer and more useful custom open models on Databricks AI Runtime. Flo first created an evaluation system of 26 AI judges (12 medical-safety, 6 micro-behavior, 8 usefulness) to turn medical standards into a training signal, generated more than 15,000 privacy-first synthetic questions, then ran supervised fine-tuning and DPO on 32 NVIDIA H100 GPUs using LLM Foundry and Compose-RL with MLflow tracking. Flo reports medical safety rising from 34% to about 61%, usefulness from 42% to 76%, and the share of responses meeting both from 16% to 47%, and estimates tailored models can cut inference costs by an order of magnitude versus frontier APIs; it plans to apply the pipeline to newer open-weight models such as Gemma, Qwen and Kimi.
Databricks Blog · 🔥 0 - Broadcom CEO Hock Tan likens AI investors to casino gamblers
Broadcom CEO Hock Tan said many AI investors "have no clue what they are investing in" and set extreme expectations only to find it is "almost like they are in a casino," speaking Monday at the DBS Asia Future Forward Conference in Singapore on a panel with Intel CEO Lip-Bu Tan, moderated by DBS CEO Tan Su Shan. He said many enterprises are "just wasting money, spending tokens" like a child spending $10 in ten minutes at an arcade, and described OpenAI's $50bn annualised revenue in September as "by itself, not that bad" for a startup that is still losing money because of heavy training spend. Both CEOs said capital could become a bigger constraint on AI than chips, memory or power; Tan said compute demand should outstrip supply through 2029, possibly into 2030, with wafer shortages easing by 2028 or 2029, and Broadcom is raising $60bn in debt to help Anthropic and other customers pay for AI chips and infrastructure.
Techmeme · 🔥 0 - Wang Yuquan: AI is no bubble, but 2029 is the risk window
Technology investor Wang Yuquan of Haiyin Capital argues in a conversation that the current AI boom is not a bubble but a decade-plus technology shift, with the real risk window around 2029, when application paradigms in AI software, AI coding and autonomous driving become clear while revenue lags behind infrastructure spending. He expects 24/7 trading, quantitative strategies and stablecoin growth to amplify market swings, hurting memory makers such as SK Hynix and South Korea hardest, while leaders like Nvidia are undervalued. He also calls embodied AI and large language models overhyped, urges Chinese labs to close their models for safety and revenue reasons, and predicts a 70-30 split favoring China outside the US.
虎嗅 AI · 🔥 0 - Intel, Broadcom CEOs warn capital could become AI's biggest bottleneck
Speaking at a panel organized by DBS Bank, Intel CEO Lip-Bu Tan and Broadcom CEO Hock Tan warned that access to capital could become the biggest constraint on the AI boom. Broadcom's Tan said the company is working through wafer and memory chip shortages but that capital will be the main limiting factor, and he likened some AI investors to casino gamblers who have little understanding of what they are betting on. Intel's Tan said Chinese frontier models such as DeepSeek, Kimi and GLM are already good enough, with low token costs and fast adoption, voiced strong support for open source, and said quantum computing could go mainstream within three to five years.
36氪快讯 · 🔥 0 - Moonshot AI in pre-IPO talks to settle 10.5% stake claim
Reuters reported on October 12 that Moonshot AI executives are negotiating a settlement with shareholder Recurrent.ai (Huanxiang Zhineng) and some of its investors, seeking to resolve before the company's Hong Kong listing a dispute that has gone to Hong Kong arbitration and dates back to 2024. Recurrent alleges that Moonshot CEO Yang Zhilin and CTO Zhang Yutao made misrepresentations, that servers holding key large language model data were moved without authorization from Recurrent to a Moonshot subsidiary in April 2023, and that other key data was deleted on Zhang's instructions in September 2024; Reuters said the letter provided no evidence and it could not independently verify the claims. The claimants seek compensation equal to 10.5% of Moonshot's original shares, which people familiar said could exceed $700 million based on the company's $30 billion valuation at its June funding, while Bloomberg reported a latest round at about $50 billion. Moonshot has confidentially filed for a Hong Kong IPO; Yang has denied the allegations, his lawyers say the case lacks legal and factual basis, and whether a settlement will be reached is unclear.
IT之家 AI · 🔥 0 - AI product naming shifts from grand cosmic narratives to everyday roles
An analysis published by Huxiu AI (author "叫我茉莉同学") traces how AI product naming has changed: early products leaned on technical mechanisms (ChatGPT, Stable Diffusion), cosmic or philosophical imagery (Tongyi Qianwen, Hunyuan, Moonshot, Midjourney, Grok) or Latin/Greek roots (Anthropic, Cohere, Mistral), answering what a product believes in rather than what it does. As technology matured and competition shifted toward scale and growth, the article argues, names moved toward defining human-AI relationships (such as Claude Cowork or Tencent's Buddy family) or toward plain, memorable labels (the author cites Nano Banana and Jaguar). Using Moonshot/Kimi and DeepSeek as examples, it says a company name speaks to capital and talent while a product name speaks to users — the name is really a claim on the product's social role. These are the author's opinions and examples, not official company statements.
虎嗅 AI · 🔥 0 - Yuandong Tian: co-writing a paper with GPT-5 made me fear job loss within 5 years
Yuandong Tian, a former Meta FAIR researcher and co-founder of Recursive Superintelligence, said on The Information Bottleneck podcast that he treated his final Meta paper — on the neural-network grokking phenomenon — as a collaboration with GPT-5, using the model to prove and check theorems and brainstorm, and estimates the workflow was roughly 6–10x faster. That experience, he said, led him to conclude his own job could be replaced within about five years, and to start a company instead: "rather than passively lose my job, I'd rather start a company heading in that direction." He was also skeptical of full autonomy, saying models asked to generate ideas on their own produce mediocre ones and that humans still frame the questions and validate results. He argued that restricting recursive self-improvement is hard because the boundary is blurry, and said Recursive will open-source its research. Claims such as the company's reported $4.65 billion valuation and over $650 million raise, and OpenAI's unreleased internal model producing 722 math manuscripts, come from the article and interview accounts.
36氪 人工智能 · 🔥 0 - Virtual quantum lab stress test: 17 AI agents, pass rates from 78% to 3%
A joint team from the National University of Singapore, Nanyang Technological University and GaugeForge released Quantum-Harbor, a simulated quantum engineering lab, plus QIQCBench, a set of 49 real-world tasks spanning measurement and characterization, control, error correction, compilation, simulation and sensing. Seventeen frontier AI agents were dropped into the physics-driven environment with instrument APIs and hidden hardware parameters, then graded under an "evidence-bound" verified-autonomy standard that checks whether conclusions are supported by data the agent actually collected. Pass rates ranged from 78% at the top down to 3%, 937 of 1,510 failures came from exhausting budget or looping, and the many-body physics category averaged only 15%.
36氪 人工智能 · 🔥 0 - GLM-5.3 joins Amazon Bedrock, showing two Chinese model go-to-market paths
An analysis published via 36Kr's Muhe Business Finance says Zhipu's GLM-5.3 went live on Amazon Bedrock during China's October National Day holiday under a usage-based revenue-share deal; Zhipu's Hong Kong shares reportedly rose more than 7% that day, and Goldman Sachs upgraded the stock to buy with a HK$1,560 target, raising its end-2026 ARR estimate for Zhipu from $2.7bn to $3.2bn. The piece contrasts this with DeepSeek, which has been available on Bedrock since March 11, 2025 via Bedrock and Custom Model Import but is described as prioritising distribution and open weights over direct monetisation, a difference it ties to Zhipu's post-IPO revenue pressure versus DeepSeek's High-Flyer backing and its reported RMB 51bn first external round in June 2026.
36氪 人工智能 · 🔥 0 - PPT Master: open-source AI workflow builds natively editable PowerPoint decks
PPT Master is an open-source Python workflow by developer Hugo He that runs inside any agent-capable AI tool (such as Claude Code or Codex CLI) and turns PDFs, DOCX files, web pages or a bare topic into a natively editable PowerPoint deck — including slide masters and layouts, native shapes and connectors, data-backed charts and tables, transitions, animations and narration from speaker notes, plus support for a user's own .pptx templates. The project bills itself as free and open source with no PPT subscription, with everything except model communication running on the user's machine to avoid platform lock-in. The author recommends a large-context model (Kimi K3 or Claude) paired with an AI image generator such as gpt-image-2 or gemini-nano-banana-2.1, and warns that a single pass will not produce a flawless deck; the page also carries sponsor promotions for Kimi, PackyCode, APIKEY.FAN, RunAPI and APIMart.
GitHub Trending · Python · 🔥 0 - Unsloth ships desktop app and Studio web UI
The open-source Unsloth project is trending on GitHub's Python list, and its README now describes three ways to use it: the native Unsloth Desktop app for Windows, macOS and Linux, the Unsloth Studio web UI, and code-based Unsloth Core. The project says it can run and train LLMs plus diffusion, embedding and audio models, naming Qwen3.8, DeepSeek-V4, Gemma 4, GLM-5.3-Flash, Kimi K3 and MiniMax-H3, and its Unsloth Start command wires local models into agents such as Claude Code and Codex. The README also claims 2x faster training with up to 70% less VRAM and no accuracy loss — figures that come from the project's own documentation.
GitHub Trending · Python · 🔥 0 - SUSE unveils private enterprise AI strategy, AI Factory with NVIDIA and Liz ops agent
SUSE announced an "enterprise private AI" strategy and an AI Factory offering with NVIDIA that combines NVIDIA AI Enterprise with SUSE's infrastructure and security governance capabilities, for AI workloads on-premises, in multi-cloud and in offline environments. SUSE says the combination can cut deployment time from months to days, though it acknowledges actual timelines depend on existing systems, data readiness and integration complexity. The release also includes Liz, a Kubernetes operations agent integrated into Rancher Prime that supports Chinese-language interaction and troubleshooting under a human-in-the-loop design; the China version is adapted to models including Qwen, DeepSeek, Kimi, Zhipu and MiniMax. SUSE also floated a possible evolution toward a "Token Factory," described as a product direction rather than a delivered capability.
InfoQ 中文 AI&大模型 · 🔥 0 - Anthropic cuts Claude Haiku 5.5 to $0.10/M tokens; OpenAI puts GPT-6 Luna on free tier
Anthropic released the low-cost lightweight Claude Haiku 5.5 on Oct 8, priced at $0.10 per million input tokens and $0.50 per million output tokens for requests under 100,000 tokens — about one-tenth of the previous Haiku 4.5 — and also cut Sonnet 5.5 cache-read pricing from $0.20 to $0.10, per an AIX Finance article. Around the same time, OpenAI said GPT-6 is fully rolled out, with GPT-6 Sol for paying users and GPT-6 Luna for free users at the same API price as Haiku 5.5. The piece compares Haiku 5.5, GPT-6 Luna and DeepSeek-V4.1-Flash on pricing and caching costs, and looks at how Chinese vendors can respond with tiering and open-weight ecosystems.
36氪 人工智能 · 🔥 0 - Ecosia drops Mistral for Chinese open-source AI models
Berlin-based search engine Ecosia, which is used by government services, is ending its work with France's Mistral and moving mainly to open-source models, including Chinese ones. Founder and CEO Christian Kroll told Politico: "We are disappointed with the quality of Mistral," saying the model now trails competitors by a full year. Reports note Ecosia had only switched from OpenAI to Mistral in May, while a Chinese-language headline claims the move to Chinese open-source models halves costs.
Techmeme · 🔥 0 - Reflection launches Beam, Mistral debuts Large 4 as Western open models push back
Reflection, the US startup billed as an "American DeepSeek", released its first open-weight model Beam: a 501B-parameter MoE with 23B activated per token for coding, reasoning and agent workloads, whose final RL run used more than 10,500 Nvidia GB300 GPUs and over 100 million rollouts. A day later Mistral launched Mistral Large 4, its largest model yet at 1.05T parameters (52B active), 1M-token context and native multimodality, trained from scratch on about 3,800 Grace Blackwell GPUs in Europe; it is in API preview with full weights due in late October. Both companies concede Chinese open models remain ahead: Beam trails GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash on DeepSWE v1.1 (44.4), Terminal Bench v2.1 (80.1) and HLE (36.2), so Reflection is pitching intelligence per unit of inference compute instead.
36氪 人工智能 · 🔥 0
Experience and discussion from the community
Share my Kimi experienceAsk about Kimi
Nobody has shared their experience with Kimi yet.