All news

336Kr · AI23:03Product release

AI score69

DeepSeek launches AI-native search engine, API billed at 5 yuan per 1,000 requests

DeepSeek announced the launch of its AI-native general search engine; the API supports web search and crawling, billed at 5 yuan per 1,000 successful requests, while the V4.1 Flash model was updated to version V4.1-Flash-1017. Official documentation states that the engine indexes hundreds of billions of web pages worldwide, supports multilingual search, and filters SEO cheating, GEO manipulation, and low-quality AIGC content. The article notes that in testing, the search API provides rerank_score and authority scores, which it sees as intended to solve the trust problem of agents acquiring online information.

#Product updates#Search#English text is being prepared.

Hheise online23:00Product release

AI score75

Anthropic launches proactive AI agent Claude Duet

Anthropic releases proactive AI agent Claude Duet, based on Claude Opus 5.5, running in a virtual machine with an independent browser, able to monitor email, calendars, and business software on its own and handle tasks in the background. The service is first available in preview to Pro and Max subscribers in the United States, requires joining a waitlist, and can be used via the iOS and Android apps and claude.ai. Duet connects to more than 6000 connectors and plugins, including Gmail, Slack, QuickBooks, Stripe, and others; actions such as payments, signing documents, and deleting files require explicit user confirmation, while securities trading and transfers between accounts are completely prohibited. Anthropic calls it an early version that may misread messages or misjudge urgency, and background checks consume existing Claude plan quotas.

Same story · 44 reports

#Product updates#Agents#Anthropic

Ggihyo.jp23:00Product release

AI score55

Google publicly releases complete 32-bit Coral NPU design and launches Coral Compiler v1.0

On October 20, 2026, Google Research publicly released the complete 32-bit Coral NPU design, including the matrix operation unit, and simultaneously released Coral Compiler v1.0, which handles model conversion and optimization. The NPU is based on the 32-bit RISC-V instruction set and targets low-power inference for wearable and IoT devices. The design data (RTL) has been made public on GitHub, and the compiler supports JAX, PyTorch, and TensorFlow models. On the same day, Synaptics released the Coralboard SR development board using the Coral NPU and began general sales of the Coralboard SL on DigiKey; the latter is priced at USD 99.79 on the U.S. site.

#Product updates#English text is being prepared.#Open-source ecosystem

TThe New Stack23:00Product release

AI score52

Microsoft launches three MAI speech models on Foundry

Microsoft launched three MAI speech models on Microsoft Foundry, including the streaming transcription model MAI-Transcribe-2-Streaming, the multilingual text-to-speech model MAI-Voice-2.1, and MAI-Voice-2.1-Flash for high-concurrency scenarios. According to Naomi Moneypenny, Senior Director of Product Development at Microsoft, streaming transcription produces an initial hypothesis within a few hundred milliseconds of receiving audio and progressively corrects it as context accumulates; the model ranks first on the Artificial Analysis leaderboard for both intermediate results and final transcription accuracy. Practitioners from the speech agent company Phonely caution that speech input and output are not the difficult part; the real challenge lies in pause handling to determine whether the caller has finished speaking, and latency and turn-taking need to be optimized in tandem.

#Model releases#Speech#Microsoft

SSina Tech22:55Product release

AI score66

iFlytek releases SparkAudio 1.0 speech large model and Robot Super Brain Platform V3.0

At the main forum of the 2026 iFlytek Global 1024 Developer Festival, Chairman Liu Qingfeng released the SparkAudio 1.0 speech foundation large model, multimodal interaction technology and Spark embodied multimodal interaction technology, and launched the Robot Super Brain Platform V3.0. SparkAudio 1.0 covers speech recognition, simultaneous interpretation, audio understanding, emotion recognition, real-time speaker separation and environmental sound understanding; related speech recognition tests show an overall effect improvement of 20%, a 15% reduction in response time and a 60% cost reduction, and it has been launched in products such as iFlytek Input Method 16.0, iFlytek Hearing and iFlytek Smart Office Book. The event also released the iFlytek AI Recording Ring, which transcribes 2 hours of recording in just 2 minutes, with comprehensive recognition accuracy exceeding 96%.

Same story · 2 reports

#Product updates#Speech#Multimodal

HHugging Face22:01First-partyProduct release

AI score52

Hugging Face launches a Space to check whether models can run on AWS Trainium

Hugging Face releases the torch-neuronx model support Space, which lets users view, by download count or trend, whether models on the Hub can run on AWS Trainium, and reports the number of sharded cores, compile-inclusive latency, and peak memory. Currently it only tests Transformers text-generation and image-text models, and only models that fit on 4 Trainium2 devices (up to 16 cores); each model generates 32 tokens for validation, comparing against CPU results in single-core mode, and in sharded mode a small model judges whether the English is fluent. Results come from the public dataset aws-neuron/torch-neuronx-model-support, and the Space also provides run scripts and a NEFF cache command that skips compilation.

#Tutorials / practice#English text is being prepared.#Open-source ecosystem

MModelBest OpenBMB21:39Product release

AI score65

ModelBest OpenBMB releases open-source AI digital employee platform StaffDeck 1.0.0-preview

ModelBest OpenBMB releases the open-source platform StaffDeck 1.0.0-preview, used to build, execute, and operate AI digital employees. This version brings modular Harness v3, a CLI and API that connect agent workflows to the terminal, asynchronous execution with traceable task status and resumable SOPs, and a human-in-the-loop mechanism that can hand over tasks and resume execution. The official release also provides links to the GitHub repository, installation, and CLI guides.

#Product updates#Agents#Open-source ecosystem

CCaixin19:59Product release

AI score58

DeepSeek open-sources inference framework DeepInfer, compatible with NVIDIA GPUs and Huawei Ascend 950

On October 20, DeepSeek announced the open-sourcing of its internal large-scale cloud service inference framework DeepInfer, aiming to improve model inference efficiency and reduce deployment costs, compatible with NVIDIA GPUs and Huawei Ascend 950 chips. The framework has been used in the online service of the DeepSeek V4.1 Flash model, with a single cluster processing an average of 155 trillion Tokens per day, and revenue of about 38 US dollars per GPU hour calculated at off-peak electricity prices. This issue of the GPT Weekly also mentions news such as Anthropic launching the personal AI assistant Claude Duet and MiniMax breaking its IPO price in the Hong Kong stock market.

#Industry news#Open-source ecosystem#Reasoning

NNumerama19:59Product release

AI score66

1Password Launches for Agents, Letting AI Agents Log Into Accounts Without Touching Passwords

1Password has released 1Password for Agents, allowing AI agents to log into user accounts, but the password is filled in by 1Password itself and immediately withdrawn after a successful login, so the agent cannot read or reuse it. The first batch of integrations is Meta's Muse and SpaceXSI's Grok Bot; integrations such as Claude Duet, ChatGPT Work, dots, and Perplexity have not yet been given a timeline. The feature is included in personal and family subscriptions and currently only supports accounts hosted on 1Password.com; 1Password.eu and 1Password.ca accounts are not yet supported.

Same story · 13 reports

#Product updates#Agents#English text is being prepared.

PPhoenix Technology19:52Product release

AI score71

OpenAI opens the Decisions API to all customers, with response times in some scenarios shortened to 100 milliseconds

On October 22, OpenAI officially opened the Decisions API to all customers. The interface optimizes inference speed and cost based on the reasoning model GPT-6 Luna, and in some scenarios shortens response time from 1 to 3 seconds to 100 to 180 milliseconds. The interface targets tasks with a clear answer range; developers can preset questions and optional answers, and the model outputs probabilities, options, or scores, for use in content classification, request routing, and deciding the next action of an AI agent, processing up to 400 questions per request. Enterprises such as Datadog, Ramp, and Speak have used it to replace some large-model calls; among them, Datadog's P95 latency for assessing agent command risk is about 40 times faster than available frontier models.

Same story · 7 reports

#Product updates#Reasoning#Agents

KKling AI17:01First-partyProduct release

AI score67

Kling 4.0 is now live on idea.ai

Kling AI announced that Kling 4.0 is now live on idea.ai. According to idea.ai's release notes, Kling 4.0 supports generating 30-second scenes from a single prompt, keeps characters consistent from shot to shot, and can switch between live-action, anime, and 3D styles, outputting cinematic-quality video.

Quoted postidea.ai · @officialidea_ai

Kling 4.0 is now live in https://t.co/dY2XctHySY 🎬 Generate 30-second scenes from a single prompt. Consistent characters from shot to shot. Switch styles from live action to anime to 3D. Create videos in cinematic quality with Kling 4.0 in https://t.co/dY2XctHySY https://t.co/gRV1ZuGgXo

Same story · 2 reports

#Product updates#Video#English text is being prepared.

KKling AI17:01First-partyProduct release

AI score62

Kling responds to Kling 4.0 launch on invideo: 30-second single generation, 4K output support

The official Kling account reposted invideo's announcement, confirming that Kling 4.0 is now live on invideo, with single generation producing 30-second clips, supporting up to 10 keyframes, 15 references, and 4K output. Kling's caption reads "consistency is key, and so it is when scoring," accompanied by the #Kling4 and #KlingAI hashtags.

Quoted postInvideo · @invideoOfficial

Kling 4.0 is now live on invideo. 30 second clips in a single generation, up to 10 keyframes, 15 references, and 4K output. https://t.co/eyhmPKGNnb

#Industry news#Video#OpenAI

SSemiconductor Industry Observer16:58Product release

AI score67

MSI releases EdgeMesa N AI+ mini PC, 1.3L chassis runs 120B large model locally

MSI has released the EdgeMesa N AI+ mini PC, with a volume of 1.3 liters, equipped with the NVIDIA RTX Spark N1X superchip and up to 128GB of LPDDR5X unified memory, with running 120B-parameter large models locally as its primary design goal. The chip integrates a 20-core Grace CPU and a Blackwell GPU with 6144 CUDA cores; the unified memory pool lets model weights and KV cache share the same physical memory. Referring to measured results from the DGX Spark on the same platform, generation speed for 120B-class models is about 15 to 59 tokens/s, limited by 273GB/s memory bandwidth; official pricing has not yet been announced, and the article estimates it may fall in the 30,000 to 40,000 RMB range.

#Product updates#English text is being prepared.#English text is being prepared.

VEnglish text is being prepared.16:58Product release

AI score58

XPeng IRON upgrade: second-generation VLA deployed on-device, re-infers every 0.2 seconds and predicts actions for the next 6 seconds

XPeng Motors announced the latest progress on its humanoid robot IRON. The upgraded second-generation VLA large model has been deployed on the robot's device side, enabling it to re-infer every 0.2 seconds based on real-time visuals and predict action changes within the next 6 seconds. The demonstration covered scenarios such as handshakes, fist bumps, and high-fives; when the interacting person temporarily changes gestures or position, IRON can adjust its actions and switch hands to coordinate. In terms of hardware, IRON is equipped with three self-developed Turing AI chips, with effective computing power up to 2250TOPS, a dexterous hand with 22 degrees of freedom per hand, and a model architecture combining VLT, VLA, and VLM.

#Product updates#Embodied AI#English text is being prepared.

Hheise online16:58Product release

AI score56

AWS Continuum officially launches code vulnerability detection and remediation capabilities

AWS has officially made available the code vulnerability detection and remediation feature in its security service Continuum; this feature verifies whether scan results can actually be exploited, prioritizes them by business impact, and provides remediation recommendations. Continuum links code repositories with AWS environment information into a context graph to determine whether an attacker can reach a vulnerability and what resources could be further accessed, and provides PoC code and reproduction steps when reproducing vulnerabilities in an isolated sandbox. Remediation is submitted as a Pull Request to the associated GitHub repository and is applied only after manual review; it is currently available only in the US East (N. Virginia) and US West (Oregon) regions.

Same story · 2 reports

#Product updates#Safety / alignment#Agents

Hheise online16:58Product release

AI score62

GitHub Package Firewall opens public preview, blocking malicious npm packages

GitHub is opening Package Firewall as a public preview to Enterprise Cloud customers, checking package queries and downloads against centralized security policies; currently only npm packages are supported. Administrators can set a waiting period after a new version is released, block based on malware and vulnerability information from the GitHub Advisory Database, and prevent downloads when install scripts such as preinstall or postinstall appear for the first time in a new version or when SLSA provenance that previous versions had is missing. The firewall sits between clients and upstream package sources, so coding agents and CI/CD hosts must route their requests through the service for it to take effect; it is free during the preview, and general availability will be billed by usage. GitHub has not announced the general availability date or specific pricing.

#Product updates#Safety / alignment#GitHub

Ggihyo.jp16:57Product release

AI score71

Meta's personal AI agent Muse launches in Japan, its first market outside North America

Meta launched its personal AI agent Muse in Japan on October 23, available on iOS, Android, and muse.ai; Japan is the first overseas market following the initial launch in North America in September this year. Muse can make plans and execute tasks on the user's behalf, continuing to work even after the app is closed, and it requests user confirmation before actions such as sending emails or purchasing products; Meta is also advancing connector integrations with Japanese local services such as Mercari, Money Forward ME, and Ticket Pia, as well as Gmail, Google Calendar, Instagram, Stripe, Shopify, and Trip.com. Meta says it runs in an isolated Muse Secure VM, with network access managed by a separate Sentinel agent; basic features are free, while the Power plan costs $20/month and the Maximum plan costs $100/month.

Same story · 2 reports

#Product updates#Agents#Meta

LLeiphone16:57Product release

AI score55

Meituan Waimai releases "Miaoshou Waimai Edition · Kangaroo Manager" AI store manager and announces free provision to restaurant merchants nationwide

At the Meituan Waimai Industry Conference on October 22, Meituan released "Miaoshou Waimai Edition · Kangaroo Manager," and Meituan Waimai General Manager Xue Bing announced that a "AI store manager" that understands food delivery operations will be provided free of charge to restaurant merchants nationwide. The product covers scenarios such as business diagnosis, store inspection, product operations, and new store operations, and can combine Meituan and merchant data to provide analytical suggestions and assist in execution. Tens of thousands of small and medium-sized merchants have already tried it. After a chicken clay pot restaurant in Yuanshi County, Hebei, adjusted its menu and promotions, its weekly order volume rose more than 30% month-on-month. The product has also cooperated with more than 200 chain brands such as Guming, Yeehoo Tea, and Lanxiangzi; Yeehoo Tea's newly opened stores saw a 25% increase in first-month per-store output.

#Product updates#Agents#English text is being prepared.

LLeiphone16:57Product release

AI score59

Qwen Go! Launches "My Apps," Supporting No-Code Building of Shareable Personal Apps

"Qwen Go!" launches a new feature "My Apps." Users can build apps for life, work, and study scenarios without programming, simply by stating their needs. The generated apps support external sharing and continuous data updates, and remain usable long-term. "My Apps" connects to Alibaba Cloud's AI application production platform Meoo (Miaowu), which can automatically complete app building, database configuration, and deployment. Combined with cloud Pro capabilities, after authorization it can call data from apps such as Taobao, Damai, and Quark Drive. Currently, "My Apps" and the cloud Pro that supports operating cloud phones and cloud computers are both in the invite-only testing phase. Users can queue up to apply on the invite-only testing page within the Qwen App.

Same story · 3 reports

#Product updates#Agents#English text is being prepared.

MMarkTechPost16:51Product release

AI score59

You.com launches Inference Router, connecting 12 open-weight models and offering a 99.9% availability SLA

You.com launches Inference Router, an OpenAI-compatible managed API that routes requests across 12 open-weight models and offers at least a 99.9% availability SLA, zero data retention by default, and 100% US-based hosting. Requests first go through You.com's own inference capacity, and when it is at full capacity or fails, they switch to US-hosted partners such as Parasail, Wafer, Alibaba Cloud, Meta, and Inception, with token-based billing and no platform fee. The built-in server-side web search returns up to 20 results per query with a default of 30 rounds, and You.com self-reports that GLM-5.3 plus its search scores 86.53% on LiveNewsBench at a cost of $0.0485 per question.

#Product updates#English text is being prepared.#Open-source ecosystem