DeepSeek open-sources its internal inference framework DeepInfer
要闻
DeepSeek has open-sourced its internal inference framework DeepInfer alongside the Token encoding and decoding component deepseek-recipe, saying the former powers DeepSeek V4.1 Flash with throughput of 155 trillion tokens per day in a single cluster.
DeepSeek has open-sourced its internal inference framework DeepInfer and the Token encoding and decoding component deepseek-recipe. The open-source version of DeepInfer uses the Apache 2.0 license and supports NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. According to the company, the DeepSeek V4.1 Flash online service runs on this framework, with single-cluster throughput of 155 trillion tokens per day and revenue of US$38 per card per hour calculated at off-peak pricing.
DeepInfer supports distributed inference, DSpark speculative decoding, SWA Bounded Replay, and structured output. DeepSeek says it will also continue to open-source inference code for future models and subsequent optimizations through DeepInfer. The other component, deepseek-recipe, handles encoding and decoding between three API protocols—Chat Completions, Responses, and Messages—and model tokens. According to the company, the component fully aligns with the official API’s processing logic.
ChatGPT introduces locked chats with PIN protection
要闻
OpenAI is gradually rolling out locked chats to ChatGPT users on Free, Go, Plus, Pro, Business, Enterprise, and Edu plans, with access protected by a 4-digit account PIN. Locking does not provide end-to-end encryption.
OpenAI is making Locked Chats available in stages in ChatGPT across the Free, Go, Plus, Pro, Business, Enterprise, and Edu plans. According to its official explanation, conversations moved to a separate area no longer appear in regular history, pinned chats, or regular search, and accessing them requires a 4-digit account PIN. Supported mobile devices also allow Face ID, Touch ID, or the device’s unlock method. When memory is enabled, new messages in locked chats can create memories for use in other locked chats, but do not update the memories used by regular chats. OpenAI states that locking does not provide end-to-end encryption or change chat retention periods or the “Improve the model for everyone” setting. Resetting the PIN permanently deletes all locked chats, files and other content protected only by the original PIN, and memories that exist only in locked chats.
Day 15 of the Codex improvement plan: Tibo resets usage limits
开发生态
OpenAI’s Tibo announced a usage reset on day 15 of an ongoing improvement plan. He had previously pledged that, for the next 28 days, each day would bring either one concrete improvement useful to most Codex and Work users or a full usage reset.
OpenAI’s Tibo posted on day 15 of the ongoing improvement plan to announce a usage reset and said that improvement work was continuing. The reset represents one of the two forms of daily delivery he had previously pledged: for the next 28 days, each day would bring either one concrete improvement useful to most Codex and Work users or a full usage reset. The commitment calls for an improvement or a usage reset each day, rather than both every day.
OpenAI fixes quota allocation errors affecting around 40,000 accounts and provides compensation
开发生态
OpenAI said it has corrected a mismatch between allowance calculations and selected plans affecting approximately 40,000 accounts and is compensating affected users with $500 in credits. The compensation is separate from weekly allowances and expires on January 20, 2027.
OpenAI said the allowance mismatch affecting approximately 40,000 accounts was corrected on October 16; the source does not specify the year of that date. Some users had reported over the weekend that their allowances were being consumed faster than expected. In an email, the support team explained that weekly allowances had not been adjusted for some users who moved from Pro 200 to lower-priced subscription plans. When the correction was applied, usage already accumulated that week was counted against the new allowance, reducing the remaining balance. Affected accounts are receiving compensation worth $500, totaling 12,500 credits. These credits are accounted for separately from weekly allowances and expire on January 20, 2027.
OpenAI has added an optional 12-hour prompt cache for GPT-5.6 and later models in the Responses API. Developers can combine it with the default 30-minute cache in a single request; both durations have the same read rate but different write rates.
OpenAI has made a 12-hour prompt caching option available for GPT-5.6 and later models in the Responses API, while keeping the default at 30 minutes. According to the company, a single request can assign shared instructions, tools, and context to a 12-hour cache and newer conversation content to a 30-minute cache, allowing requests spaced hours apart to reuse shared content at cached input rates. Cache writes for the 12-hour and 30-minute options cost 2.5 times and 1.25 times the standard input rate, respectively, while read rates do not vary by cache duration. Cache reuse requires two conditions: an exact match at the beginning of the prompt and fulfillment of the model’s minimum cacheable length.
Google open-sources TorchTPU, a native PyTorch software stack
开发生态
Google announced at PyTorch Conference 2026 that TorchTPU is now open source on GitHub, making TPUs a native execution option for PyTorch. It also outlined model-library support, inference-engine integrations, and the release of llm-d 1.0.
Google announced at PyTorch Conference 2026 that TorchTPU, its native PyTorch software stack for Google TPUs, is now open source on GitHub. Google jointly developed the stack with Meta’s PyTorch team. According to Google, existing PyTorch workloads can run on TPUs with minimal code changes, using familiar APIs and eager mode for debugging before applying torch.compile to optimize TPU utilization and reduce memory overhead. For model libraries, Transformers v5 supports deploying models to TPUs with device_map="tpu". Diffusers runs natively on TPUs and supports eager mode, torch.compile, and tensor parallelism. Google also built CI/CD pipelines for continuous nightly testing of the Hugging Face Transformers suite on TPU infrastructure.
For inference deployment, Google integrated TorchTPU with open-source inference engines including vLLM and SGLang, alongside Kubernetes-native fleet routing for TPUs and GPUs. The native PyTorch backend for SGLang, developed with RadixArk, preserves upstream worker models, schedulers, and API servers. SGLang uses RadixAttention to reuse KV cache across requests, avoiding recomputation of identical prompt prefixes. Google also released llm-d 1.0 with Red Hat, IBM, and a community comprising more than 500 contributors and 25 organizations. A CNCF Sandbox project, llm-d uses prefix-cache-aware routing to send requests to replicas that already hold the relevant context, while separating compute-heavy prefill from memory-heavy decode. Version 1.0 provides multi-tenant flow control and router high availability.
腾讯元宝 search opens to external users with three integration options
开发生态
Tencent has opened Yuanbao Search through Tencent Cloud, offering three integration methods—Skill, MCP, and API—and a monthly allowance of 2000 free calls. Tencent says the service covers 157 authoritative sources across 13 industry verticals.
Tencent has officially made Yuanbao Search available externally through Tencent Cloud. On retrieval performance, the company reports an average underlying index retrieval time of 8 milliseconds and a recall rate of 94%. According to Tencent, Yuanbao Search has built an index on the scale of 100 billion entries and connects to 157 authoritative sources across 13 industry verticals, including finance, healthcare, law, and travel, with minute-level updates supported for content such as policies, news, weather, and market data. The three integration methods are Skill, MCP, and API, with a monthly allowance of 2000 free calls available upon integration.
ElevenLabs releases two Eleven Realtime voice models
模型发布
ElevenLabs has released two Eleven Realtime voice models, both supporting more than 90 languages and over 17,000 voices. According to the company, developers can access them through ElevenAgents or an interface compatible with OpenAI’s Realtime API.
ElevenLabs has launched Realtime v1 Max and v1 Mini, Speech to Speech models that integrate the capabilities of a complete voice Agent. The company says both support more than 90 languages and can use over 17,000 voices from its voice library. Artificial Analysis independently evaluated both models. On Big Bench Audio, which contains 1,000 spoken reasoning questions, Max and Mini achieved accuracy rates of 98.8% and 75.9%, respectively, with average times to start speaking of 0.57 seconds and 0.62 seconds. On Full-Duplex-Bench, which evaluates pauses, turn-taking, interruptions, and backchannels, Mini scored 95.8% and Max scored 94.1%.
Access is available through ElevenAgents or the separate Realtime API. According to the company, the former provides phone, web, WhatsApp, and CCaaS channels, along with workflows, a knowledge base, guardrails, testing, and monitoring. The latter follows OpenAI Realtime API specifications and works with the existing OpenAI SDK, but supports only WebSocket, so existing WebRTC integrations must migrate first. The standalone API does not include ElevenAgents platform features. In Speech Agent Arena, where users compare models through blind live conversations, Max and Mini scored 998 and 988, respectively; their completion rates on tool-use tasks were 87.1% and 60.3%. Based on these results, ElevenLabs recommends Max for use cases involving frequent tool calls.
Mistral AI releases the Voxtral 3 voice model family and a voice Agent SDK
模型发布
Mistral AI has introduced the Voxtral 3 speech model family and a voice Agent SDK, combining transcription, reasoning, tool calling, and speech generation in one real-time workflow, with access through an EU-hosted API or deployment on an organization’s own infrastructure.
Mistral AI has released the Voxtral 3 speech model family and voice Agent SDK, now available through its EU-hosted API or for deployment on an organization’s own infrastructure. The SDK integrates listening, reasoning, tool calling, and speaking into a real-time streaming pipeline. According to the company, Agents can retrieve information, access business systems, and complete multi-step tasks during phone or web conversations; organizations can control where audio is processed, how data is handled, and whether conversations are retained. The company also says Pro Voice can use 30 minutes of audio to reproduce a speaker’s vocal characteristics, emotional range, and speaking style across languages.
Transcription supports 50 languages, including the EU’s 24 official languages. According to Mistral AI, Voxtral Small Transcribe 3 supports diarization, context biasing, and end-of-turn detection; the Mini transcription variant has 4B parameters and costs $0.0035 per minute. Both transcription models can accept keywords and free-form text to improve recognition of specialized terminology without fine-tuning. The speech generation model, Voxtral Mini TTS 3, supports more than 35 languages and more than 50 dialects and, according to the company, can adjust emotion, pace, and emphasis to the conversational context.
OpenAI introduces recurring tasks in ChatGPT that multiple users can manage together
产品应用
OpenAI has introduced team tasks for ChatGPT Business and Enterprise workspaces, allowing colleagues to review the results of recurring engineering work and adjust future instructions together. According to OpenAI, tasks run in the cloud without requiring ChatGPT to remain open.
OpenAI has introduced team tasks in ChatGPT for Business and Enterprise workspaces; the source does not specify a launch date. According to the official help documentation, workspace permissions control task creation and management, and colleagues can jointly review results and change instructions for future tasks. OpenAI says tasks can run on a schedule or be triggered by supported app events, with execution taking place in the cloud without requiring ChatGPT to remain open. Examples of engineering work include weekly summaries of released work and blocked items.
Google launches new Gemini tools for small business owners
产品应用
Google has launched a Gemini business landing page and tools for small business owners, but the supplied source lists no launch date, tool names or pricing, leaving specific features and availability in Japan unconfirmed.
Google has launched a Gemini business landing page and tools for small business owners, although the supplied source does not specify a launch date. The information comes from Google's official blog, but the readable news content is limited to the headline. The remaining material consists of webpage styling code and provides no descriptions of the tools' features, usage requirements, pricing or version numbers. The material also does not identify supported countries or languages, or state whether users in Japan can access the tools, so availability in Japan cannot be confirmed from this source.
Google brings Gemini personalization features to Android in-car systems
产品应用
Google announced that it is bringing Personal Intelligence to Gemini in cars globally to provide personalized features. The supplied material does not disclose a launch date, the number of countries covered or supported vehicles, so it does not establish when users in Japan will have access.
Google announced Personal Intelligence for Gemini in cars globally, but the supplied material gives no effective date. The announcement’s headline positions the update as personalization for in-car Gemini, while the supplied body consists mainly of webpage styling code and provides no feature description, usage requirements or rollout schedule. It also lists no supported countries, languages, vehicles or pricing, leaving the material insufficient to confirm whether Japan is included or when users will be able to access it.
千问 App launches a shopping companion feature for asking about products by voice while sharing your screen
产品应用
Qianwen App has introduced “Qianwen Shopping Companion” for the Double 11 shopping season, allowing users to ask about products by voice while browsing shopping apps during a screen-sharing call. Access is rolling out gradually and requires the latest app version; users without an entry point must continue to wait.
Qianwen App launched “Qianwen Shopping Companion” during the Double 11 shopping season and is gradually opening access to users. To use it, users must update the app to the latest version, then access screen sharing through the call icon in the upper-right corner. Users who do not yet see the entry point cannot currently use the feature. Once a call and screen sharing are active, users can browse shopping apps and ask by voice about product reviews, specifications, prices and parameters. According to the feature description, Qianwen answers based on the current screen content and can also organize negative feedback from product reviews, calculate price differences between different specifications and remember users’ previous preferences.
Musk: X adjusts two types of models used for timeline recommendations
产品应用
Musk said X has adjusted two types of models for timeline recommendations, targeting predictions of time spent on content and prioritizing retrieval of posts users will read in full. He also claimed that user-active seconds are growing exponentially, attaching a chart comparing data from 2023 through 2026.
Musk said X has introduced a new AI/SI ranking model and trained its retrieval model to adjust timeline recommendations; the source does not specify when the changes took effect. According to his explanation, the ranking model more accurately predicts how long users will spend on content, while the retrieval model prioritizes posts users will read from start to finish. On usage, Musk claimed that X’s user-active seconds are growing exponentially and setting records almost every day. His attached Daily UAS chart compares data from 2023 through 2026, with the 2026 curve above those of the other years in its later portion.
Anthropic designs scorpion venom-binding proteins, with 17 passing binding tests
技术与洞察
Anthropic reported an experiment using an internal Claude research model to design scorpion-toxin-binding proteins: 17 of 211 laboratory-tested designs met its binding criteria. The team released all designs and data, but has not confirmed whether the proteins can neutralize scorpion venom.
Anthropic released a study on Claude-assisted design of scorpion-toxin-binding proteins, reporting that 17 of 211 tested designs met its binding criteria and making all designs and data available. The source does not specify the study’s publication date. According to the company, an internal research model running within Claude Science worked mainly autonomously for 64 hours, receiving an initial prompt developed by the team with Citrini Research and four follow-up instructions. It generated 318,484 protein designs and selected 180 for testing; the research team added another 31 with Claude’s help. Adaptyv Bio and Twist Bioscience conducted the laboratory binding tests, and each qualifying design bound to one of four scorpion neurotoxins known to be dangerous to humans. The company said it has not confirmed whether these binders can neutralize scorpion venom in a cellular environment and therefore does not claim to have produced an antivenom.
OpenAI acquires drug development platform Orchestra Bio
行业动态
OpenAI has acquired drug development platform Orchestra Bio and plans to integrate its team and technology with the Rosalind models and tools for life sciences research, building Agents that connect scientific discovery with work to advance medicines.
OpenAI has completed its acquisition of drug development platform Orchestra Bio and plans to integrate its team and technology with the Rosalind models and tools. According to the company, their combined capabilities will be used to build Agents that connect scientific discovery with work to advance medicines. In terms of their respective roles, GPT-Rosalind helps scientists reason through biology and chemistry problems, while Rosalind Workbench connects scientific tools and data to support the analysis of results and experiment planning. Orchestra Bio, meanwhile, enables researchers to delegate project monitoring and tracking to AI Agents. The Agents identify affected activities and schedules, consolidate information, and compare response options, while researchers retain responsibility for deciding the next action.
Former head of OpenAI’s Preparedness team announces return to OpenAI
行业动态
MIT professor Aleksander Madry, who previously led OpenAI’s Preparedness team, announced that he will return to OpenAI to work on AI safety and security. Sam Altman confirmed that Madry has accepted the invitation to return.
MIT professor Aleksander Madry announced that he will return to OpenAI to help address the safety and security challenges posed by increasingly capable AI; the source does not specify an effective date for his return. He previously led OpenAI’s Preparedness team. Discussing his return, Madry said AI development has entered a different phase and that addressing these problems must become a central priority. He also reframed what he had considered a departure as something that now feels more like a sabbatical. Sam Altman said Madry had accepted OpenAI’s invitation to return and described the challenges associated with superintelligence as exceptionally difficult.
Reuters: OpenAI in talks to raise around $30 billion
行业动态
OpenAI is discussing roughly $30 billion in funding and seeking investor commitments before Anthropic’s expected IPO in mid-November, Reuters reported, citing people familiar with the matter. The round could value OpenAI at more than $1.4 trillion.
OpenAI is pursuing a new funding round of roughly $30 billion, aiming to secure investor commitments before Anthropic’s anticipated IPO, Reuters reported, citing people familiar with the matter. Anthropic is expected to go public in mid-November; the report does not specify the year. According to the sources, the two companies are increasingly targeting overlapping groups of investors, and OpenAI wants to secure funding first while using recent revenue growth to attract investment. The round could value OpenAI at more than $1.4 trillion, with existing investors including SoftBank and Abu Dhabi-based MGX discussing participation. By comparison, OpenAI was valued at $852 billion after previously completing a $122 billion funding round in March; the report likewise does not specify the year for that month.
According to The Wall Street Journal, Anthropic’s draft S-1 shows first-half revenue rising from $1.1 billion a year earlier to $16.3 billion, alongside a cloud capacity shortage and potential spending of $750 billion over the coming years to secure additional computing resources.
According to The Wall Street Journal, Anthropic disclosed revenue growth, computing expenses and a cloud capacity shortage in its draft S-1 IPO filing. The company recorded a first-half net loss of $80.7 billion, including $72 billion in noncash accounting adjustments related to convertible notes and financing instruments; revenue for the period was $16.3 billion, compared with $1.1 billion a year earlier. Anthropic estimates that spending to secure additional computing resources over the coming years could reach $750 billion, including deals still under negotiation. As AWS, Google and Microsoft increasingly view it as a competitor, the company is leasing data centers and working with third-party operators, but is concerned that these operators have limited operating histories and may be less reliable than major cloud providers.
Google: Argon usage is roughly three times that of its predecessor at the same stage
行业动态
Alphabet CEO Sundar Pichai said Google’s Gemini 4 Argon processed nearly 25 trillion tokens in its first 72 hours, about 3 times Gemini 3.0’s usage over the same launch period, while serving capacity remained constrained.
Following Google’s release of Gemini 4 Argon, token usage reached nearly 25 trillion in the first 72 hours, Alphabet CEO Sundar Pichai disclosed in a Big Technology Podcast interview; the source does not specify the release date. According to Pichai, that was about 3 times Gemini 3.0’s usage in its first 72 hours, and serving capacity is currently constrained. He said Google would continue investing in frontier AI research rather than focus solely on lightweight models such as Flash, with its cloud business redirecting some compute and forgoing short-term revenue to support that work. He also said internal developers using Argon could tackle problems that previous models could not solve, although some coding tasks still needed improvement. On the deployment timeline, he said Argon’s capabilities prompted Google to spend longer on safety testing and careful deployment.
