Boss lens: a clearly priced everyday work model. Ask the tech team to compare it with the mid-tier model you already use on speed, accuracy and unit cost; if Orchestrator’s 24% quota saving holds on your own tasks, the same workload may cost less (inference, re-test required).
AI NEWS
Today’s AI News
Chinese and English are the same day. This page lists every item, each with the original link, the author, and the engagement.
Latest day
November 13, 2026
11-12 07:50 – 11-13 07:50 (UTC+8)
Top 3
Boss lens: notes are included in the plan, and in-meeting docs and boards have a free window through early February—worth a trial on internal standups. External guest meetings stay off by default on Enterprise, so an owner must turn that on, and everyone on the call should be told they will be transcribed.
Boss lens: short-video teams can test whether One-Take’s one-minute-in, one-minute-out speed holds in production; teams already on Qianfan should have engineering check whether the “over 30% cheaper inference” shows up on their own bill (inference).
All
Google launches Gemini 4 Hydrogen at $1.50/$7.50 per million input/output tokens, scoring 51 on the Intelligence Index at default and 54 at xhigh, and says it is 2.5x faster than 3.8 Flash
Boss lens: a clearly priced everyday work model. Ask the tech team to compare it with the mid-tier model you already use on speed, accuracy and unit cost; if Orchestrator’s 24% quota saving holds on your own tasks, the same workload may cost less (inference, re-test required).
CollapseGoogle launches Gemini 4 Hydrogen at $1.50/$7.50 per million input/output tokens, scoring 51 on the Intelligence Index at default and 54 at xhigh, and says it is 2.5x faster than 3.8 Flash
Google on Nov 12 launched Gemini 4 Hydrogen, a model for everyday coding and agentic execution meant to pair with longer-horizon Carbon and Argon. Official pricing is a flat $1.50 per million input tokens and $7.50 per million output tokens across a 1 million-token context; cached input is discounted 95% to $0.075. Google said the default high setting scores 51 on the Artificial Analysis Intelligence Index, ahead of Argon’s 50, and reaches 54 at the new xhigh setting. Versus Gemini 3.8 Flash, it finishes tasks 2.5x faster at the default setting and lifts Terminal-Bench 4.1 from 18.2% to 57%. Google Antigravity also adds Orchestrator, which switches between Carbon and Hydrogen in one session; Google said it reached near-Carbon success on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster on average. Hydrogen is rolling out in the Gemini API, the Gemini app, Search, Gemini for business, and Antigravity for paid subscribers, with free app users getting it via automatic model selection.
Anthropic launches Claude Meetings: Claude can join Google Meet, draft docs and whiteboards live, in beta on Team and Enterprise, with in-meeting docs and boards free through Feb 10, 2027
Boss lens: notes are included in the plan, and in-meeting docs and boards have a free window through early February—worth a trial on internal standups. External guest meetings stay off by default on Enterprise, so an owner must turn that on, and everyone on the call should be told they will be transcribed.
CollapseAnthropic launches Claude Meetings: Claude can join Google Meet, draft docs and whiteboards live, in beta on Team and Enterprise, with in-meeting docs and boards free through Feb 10, 2027
Anthropic’s official account announced in the early hours of Nov 13 that Claude Meetings is starting to roll out in beta on Team and Enterprise plans. Claude joins Google Meet as a visible participant, taking notes, answering questions, and building docs, whiteboards or slides while people talk; afterward, notes, action items, the transcript and anything it built sit on the meeting page in Claude. The help center says it saves a transcript, summary and screen descriptions, not audio, and can stay up to eight hours. Notes, transcripts and answers that do not use connected tools are included on Team and Enterprise plans; docs and whiteboards made in a meeting are free through Feb 10, 2027; connected-tool use counts toward a member’s usage and can incur charges. Organizations with zero data retention, HIPAA-ready setups or customer-managed keys cannot use it during the beta. The official post had about 38,000 likes and 9,200 bookmarks after a metrics refresh. TechCrunch and others reported the feature on Nov 10; today’s post is the official rollout thread.
Baidu World launches Ernie Lightning and Ernie One-Take: official figures claim inference costs down over 30% and video generation 50% cheaper than peers, with a real-time mode that makes one minute of video in about one minute, both on Qianfan APIs
Boss lens: short-video teams can test whether One-Take’s one-minute-in, one-minute-out speed holds in production; teams already on Qianfan should have engineering check whether the “over 30% cheaper inference” shows up on their own bill (inference).
CollapseBaidu World launches Ernie Lightning and Ernie One-Take: official figures claim inference costs down over 30% and video generation 50% cheaper than peers, with a real-time mode that makes one minute of video in about one minute, both on Qianfan APIs
On Nov 12 at Baidu World 2026, Jia Lei, head of Baidu’s applied-model team, launched two Ernie application models—Ernie Lightning for deep thinking and Ernie One-Take for micro-expression video—and opened APIs on Baidu’s Qianfan cloud. Science and Technology Daily reported that Lightning was first built for Baidu Search, uses a sparse-attention method called BSA to compress memory during decoding, and is co-optimized with Kunlun chips; Baidu said it can run high-speed inference on a single eight-chip box and cuts inference-engine costs by more than 30%. Phoenix also cited an official claim that model ROI is 50% higher than comparable chips, with thinking and non-thinking modes. One-Take targets short-form, high-expression video, accepting up to 4,000 characters of text plus visual and audio input; Baidu said generation costs are 50% lower than peers, and a fast mode can produce one minute of video in about one minute. Neither report published a public token or per-second list price for these two models, so this item does not repeat third-party price figures.
Replit Agent can now buy supplies and run Google Ads, with the owner approving every charge; ads are available to Pro users worldwide, and Replit cites a founder who built in a weekend for under $400 an app previously quoted at $105,000
Boss lens: both buying and ads stop for your approval, so a one-person company can trial a small daily budget. The revenue figures are company-told stories—ask for dashboard or invoice evidence before treating them as a benchmark (inference).
CollapseReplit Agent can now buy supplies and run Google Ads, with the owner approving every charge; ads are available to Pro users worldwide, and Replit cites a founder who built in a weekend for under $400 an app previously quoted at $105,000
Replit said on Nov 12 that Agent can now buy business supplies through Stripe’s Link and Visa, and run Google Ads across Search, YouTube, Display and Discover. For purchases, Agent prepares the cart and checkout, then waits for the owner to approve the total—it does not get open-ended spending permission, and general browsing is disabled during payment. Commerce is starting in the US for Core and Pro users and is limited to one-time purchases. Ads are available to Pro users worldwide: Agent can draft copy, generate images, recommend daily budgets and add conversion tracking, but the user still approves budgets and launches. Replit also cited two company-provided stories: Jon Cheney was quoted $105,000 by a development shop, then built the site and admin tools on Replit in a weekend for under $400 and reached six figures in the first six weeks; Austin Eckman had paid $50,000 for an unusable first version, later ran Crafting Table from Replit, and brought in $170,000 in 30 days with about 4x return on ad spend. These are self-reported cases, not independently audited.
Volcano Engine launches Seedance 2.5 Flash, claiming 2x generation speed and about 0.99 yuan per second for 720p text- or image-to-video, roughly 34% cheaper than Seedance 2.5, with the Ark API due Nov 15
Boss lens: at about 0.99 yuan per second for 720p, Flash is priced like 2.0 and about a third cheaper than the 2.5 flagship, which suits batch drafts. Keep 1080p or 4K finals on the flagship-plus-draft path, and move volume into your own stack after the Nov 15 API opens (inference).
CollapseVolcano Engine launches Seedance 2.5 Flash, claiming 2x generation speed and about 0.99 yuan per second for 720p text- or image-to-video, roughly 34% cheaper than Seedance 2.5, with the Ark API due Nov 15
On Nov 12, ByteDance’s Volcano Engine released Seedance 2.5 Flash. Per Sina Tech, the company said quality is close to flagship Seedance 2.5, generation is twice as fast, and overall quality beats the previous flagship 2.0. Pricing matches Seedance 2.0: about 0.99 yuan per second for 720p text-to-video or image-to-video, roughly 34% cheaper than Seedance 2.5. A clip can run up to 30 seconds, and a request can take up to 10 images, five video clips and five audio clips, at 480p or 720p. The model is rolling out on Jimeng, LibTV, Updream, Creao AI and TapNow, with the Volcano Ark API planned for Nov 15. The 2.5 flagship remains the high-quality path, including a draft-then-final workflow for 1080p and 4K; Flash is aimed at frequent, batch, cost-sensitive production.
Doubao Work adds Kimi 3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash, publishes credit-consumption multipliers, and starts limited-time discounts
Boss lens: you may not need a new vendor login just to try another model. Use one workbench, pick by task, and watch multipliers. Before renewing, have ops or engineering copy the live rates and time-limited discounts into a simple comparison sheet (inference).
CollapseDoubao Work adds Kimi 3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash, publishes credit-consumption multipliers, and starts limited-time discounts
On Nov 12, ByteDance’s workplace product Doubao Work said it now includes third-party models such as Kimi 3.1 Pro, GLM 5.3 and DeepSeek V4.1 Flash, so users can pick a model per task. GeekPark and TMT Post reported that the model picker now shows a credit-consumption multiplier for each model—the lower the multiplier, the fewer platform credits the same task uses. Two limited-time offers are running: discounted credit use for featured models in set weekday and holiday hours, and lower multipliers for Doubao 2.1 Lite and DeepSeek V4.1 Flash. Coverage described Kimi 3.1 Pro as the long-context option, GLM 5.3 as stronger on Chinese, and DeepSeek V4.1 Flash as the fast, cheaper daily option, alongside the existing Doubao 2.1 family. Exact multipliers and discount rates should be read from the live product page; this item does not repeat unofficial figures.
Baidu says its token exchange Ciyuansuo signed up 100-plus suppliers in a month, with more than 50 million calls and 700 billion tokens a day; prices for the same model can differ by 2x
Boss lens: the same model name can cost twice as much and respond twice as slowly depending on the supplier, so compare vendor, price and latency, not the label alone. The terms also say the platform will not cover third-party outages—keep a backup path for critical work (inference).
CollapseBaidu says its token exchange Ciyuansuo signed up 100-plus suppliers in a month, with more than 50 million calls and 700 billion tokens a day; prices for the same model can differ by 2x
At Baidu World on Nov 12, Baidu disclosed figures for its model-exchange platform Ciyuansuo after one month live: more than 100 supplier listings, over 50 million calls a day, and more than 700 billion tokens a day. Per NetEase Tech, Robin Li criticized aggregator gateways that sell access without standing behind quality or uptime, and sometimes without verifying that a model is genuine. Ciyuansuo is positioned as a neutral marketplace: buyers purchase a supplier’s inference service at stated price and availability terms, not the model itself. As of Nov 12 the catalog showed 40-plus models, including DeepSeek, Kimi, Zhipu GLM, MiniMax and Baidu’s Ernie line. For DeepSeek-V4.1-Flash, five listings quoted output tokens from 2.96 to 6.00 yuan per million, with time-to-first-token from 951 to 2,114 milliseconds. The report also noted a Nov 11 service-terms update: Baidu does not take breach liability for third-party outages, model degradation or shutdowns, and any damages are capped at fees the user actually paid in the prior 12 months. The company has not published a take-rate, so call volume cannot be translated into platform revenue.
Cerebras posts $230.9 million in core Q3 revenue, up 70% year on year, raises full-year guidance to $940–960 million, and says it signed a cloud deal with a top-three hyperscaler and will arrive on Oracle Cloud in 2027
Boss lens: besides Nvidia, another high-speed inference vendor is talking about cloud distribution and capacity. The top-three cloud partner is still unnamed officially—do not treat press speculation as a signed, named contract when you evaluate vendors (inference).
CollapseCerebras posts $230.9 million in core Q3 revenue, up 70% year on year, raises full-year guidance to $940–960 million, and says it signed a cloud deal with a top-three hyperscaler and will arrive on Oracle Cloud in 2027
AI chip company Cerebras reported third-quarter results ended Sept 30 on Nov 12. The official release put GAAP revenue at $210.8 million, up 55% year on year, and core (non-GAAP) revenue at $230.9 million, up 70%. Core cloud and other services revenue was $144.9 million, up 272%. Full-year 2026 core revenue guidance was raised to $940–960 million, with fourth-quarter core revenue seen at about $310–330 million. Remaining performance obligations were $26.1 billion at quarter-end, with $7.8 billion in cash and short-term investments. Cerebras said it signed an important cloud deal with “a top three hyperscaler,” with more to be shared next year, and that it will be available in Oracle Cloud Infrastructure in 2027. It also said it has secured 700 megawatts of data-center capacity for delivery by the end of 2027. Some English-language reports named Microsoft as that hyperscaler; the official release does not, so this item does not.
Amplitude opens its product agent Wave to paying customers, saying DigitalOcean got 55 improvement opportunities in six weeks and PixAI saw nearly 10% higher publishing and a 3% lift in day-1 retention
Boss lens: if you already pay for Amplitude, have a product lead try Wave against one concrete goal, such as signup completion. The 10% and 3% figures are customer-reported and may not transfer to your product (inference).
CollapseAmplitude opens its product agent Wave to paying customers, saying DigitalOcean got 55 improvement opportunities in six weeks and PixAI saw nearly 10% higher publishing and a 3% lift in day-1 retention
Analytics company Amplitude said on its blog that Wave, a product agent, is now a paid offering for customers on Growth and Enterprise plans. Wave investigates behavioral data, session replays and customer feedback against a business goal, turns the evidence into documented Opportunities, and helps teams follow through and measure results. The post cites two customers: DigitalOcean product manager Haley Eidem said Wave surfaced 55 opportunities across five product areas in six weeks, and her time writing standalone PRDs went to almost zero; image community PixAI used a Wave finding to point new users to a Missions tab, then reported 76% more Missions opens, nearly 10% higher publishing, and a 3% rise in day-1 retention. Amplitude also said an internal test of more personalized suggestion cards led to roughly 160% more suggestion sends. All of these figures come from the company’s blog. The post itself does not show a date; TestingCatalog wrote about the product in the early hours of Nov 13.
Voice-AI company Retell says its annualized revenue run rate has passed $100 million, after raising about $4.6 millionUnconfirmed
Boss lens: if the $100 million run rate can be tied to contracts or financials, some voice-agent vendors are already at scale. Still request the full official release, and ask whether the figure is recurring contracted revenue or includes trials (inference).
CollapseVoice-AI company Retell says its annualized revenue run rate has passed $100 million, after raising about $4.6 million
According to a Nov 12 Forbes report and a Nov 11 Business Wire headline, voice-agent company Retell said its annualized revenue run rate has surpassed $100 million. Secondary write-ups also said the company has raised about $4.6 million and introduced Tell 1, a voice model for enterprise contact centers. TestingCatalog wrote that Retell called Tell 1 its first model built for enterprise voice agents and said it ranked first on Retell’s cited vCX benchmark. This round could not retrieve the full Business Wire release, and the Forbes page available here was only a short summary, so the $100 million figure, the $4.6 million total, and Tell 1’s training or customer details have not been checked line by line against an official full text and are marked unverified. If confirmed, it would be a high-revenue, lightly funded growth case for teams buying voice customer service.
iFlytek opens a SparkAudio 1.0 API on its developer platform, supporting 99 languages and 202 dialects and understanding speech directly rather than only transcribing first
Boss lens: if your call center or meetings involve dialects, multiple languages or emotion cues, have engineering compare this with a transcribe-then-analyze setup on latency and missed context. Wait for the platform’s list price before budgeting (inference).
CollapseiFlytek opens a SparkAudio 1.0 API on its developer platform, supporting 99 languages and 202 dialects and understanding speech directly rather than only transcribing first
On Nov 12, iFlytek said its SparkAudio 1.0 speech foundation model is available as an API on the iFlytek Open Platform. Per IT Home, the model accepts text or speech and returns text, covering transcription, multilingual translation, dialect recognition, environmental-sound understanding, speaker separation, emotion analysis and audio question answering. The company contrasted this with older cascaded pipelines that transcribe first and then send text to a language model, which it said drops tone, emotion and background sound and can stack errors and latency. SparkAudio 1.0 is described as hearing speech, ambient sound and music in one pass and inferring meaning, emotion and scene, with support for 99 languages and 202 dialects, trained on domestic compute. The IT Home report did not publish API prices, and this item does not repeat unchecked billing figures.
Tongxin Intelligent, started by Zhu Songchun, raises several hundred million yuan in seed and angel funding, plans FPGA prototype validation in early December, and aims its cloud product at Nvidia’s B300 class
Boss lens: this is still a pre-tape-out startup, and December is only a prototype check, not a chip you can order. The team mixes AGI research and chip-architecture experience; execution through production will matter more than the Nvidia comparison (inference).
CollapseTongxin Intelligent, started by Zhu Songchun, raises several hundred million yuan in seed and angel funding, plans FPGA prototype validation in early December, and aims its cloud product at Nvidia’s B300 class
On Nov 12, the Beijing News and National Business Daily reported that Tongxin Intelligent, a Beijing chip startup, had completed several hundred million yuan in seed and angel funding. Backers include the Tsinghua Alumni Seed Fund, the Beijing Institute for General Artificial Intelligence, Songlin Technology, Pingzhi Information, Darkmatter and Yuzi Zhangquan, with proceeds going to pre-tape-out architecture work. The company was founded in April 2026 by Zhu Songchun, director of BIGAI and a chair professor at Peking University and Tsinghua. NBD also said chairman and CEO Wang Haiyang was an early Loongson engineer and previously led Hygon’s fourth-generation chip architecture. The Beijing News said the chip uses a three-unit Vector, Tensor and AGI Core design and hybrid in-memory computing, with Chiplet configurations planned for cloud, edge and device; the cloud product is positioned against Nvidia’s B300 class and the device product against Thor, with FPGA prototype validation planned for early December. Reports only said “several hundred million yuan,” with no more precise figure.
Wharton’s Ethan Mollick lab open-sources Primer, which turns a teacher’s spoken lesson goals into a playable student simulation, with the paper, prompts and code released
Boss lens: teams that run internal training can have L&D try one course with the public prompts. The lab itself says it has not been evaluated for classroom outcomes or student privacy, so legal review should come before using it with staff or students (inference).
CollapseWharton’s Ethan Mollick lab open-sources Primer, which turns a teacher’s spoken lesson goals into a playable student simulation, with the paper, prompts and code released
Ethan Mollick of Wharton’s Generative AI Lab posted in the early hours of Nov 13 that the lab is releasing Primer, an open-source system in which a teacher talks through lesson goals, student level and common misconceptions; the system first produces a reviewable design document and builds a playable simulation only after the teacher approves it. The lab’s site says authoring is split among intake, design and related agents, and runtime uses mentor, practice and feedback roles, in either roleplay or teach-back mode. Primer is model-agnostic and can run on local or open-source models. The paper, prompts and code are public. The lab cautions that Primer is an experimental prototype, not evaluated as a classroom intervention or reviewed for student-privacy rules, and that institutions using it with students must do that review themselves. Mollick’s post drew about 210 likes and 180 bookmarks.
Bao Yu translates “What software engineers should do now”: rather than grinding on stacks, invest in product, marketing and human judgment that AI still cannot own
Boss lens: if you still judge engineers only on coding speed, you may miss product sense and customer judgment. The 425 bookmarks make it a useful discussion piece for the team, not a hiring policy by itself (inference).
CollapseBao Yu translates “What software engineers should do now”: rather than grinding on stacks, invest in product, marketing and human judgment that AI still cannot own
Developer Bao Yu (dotey) posted a Chinese translation in the early hours of Nov 13 of an essay titled “What software engineers should do now.” The original author argues that each new model further shrinks the value of raw coding and stack knowledge, and that engineers increasingly resemble product managers of AI—sometimes a user story prompt works better than a formal technical design. The piece admits it is painful to watch once-scarce SaaS-building skill become easier, and urges a shift toward larger problems and toward the commercial loop of product, marketing and sales. The author says they are now studying positioning, creativity, how to delight people, and psychology—areas without a standard answer or a ready test set, and therefore harder for AI to replace. Bao Yu did not name the original author or link in the post. The translation drew about 420 likes and 425 bookmarks, with bookmarks slightly above likes.
Google opens pre-orders for Gemini intelligent eyewear with Warby Parker, Gentle Monster and Samsung: Warby Parker from $395, Gentle Monster from $449, in stores Dec 10, starting in the US and South Korea
Boss lens: translation and navigation could help frequent travelers, but the product is limited to US and Korean accounts for now, so it is not a default tool for China-based teams. If anyone expenses a pair, require the privacy switches and indicator light, and do not record at a customer site without notice (inference).
CollapseGoogle opens pre-orders for Gemini intelligent eyewear with Warby Parker, Gentle Monster and Samsung: Warby Parker from $395, Gentle Monster from $449, in stores Dec 10, starting in the US and South Korea
Google said on Nov 12 that pre-orders are open for intelligent eyewear made with Warby Parker, Gentle Monster and Samsung. The official blog lists Warby Parker optical and sun styles from $395 and Gentle Monster styles from $449, with prescription lenses on most frames. The glasses weigh about 41 grams; Google said battery life is up to about nine hours, and the case can recharge them up to about seven times. Features include “Hey Gemini” or tap-to-ask help, Maps directions, real-time translation across more than 70 languages, a 12-megapixel ultra-wide camera, and Guided Vision for people with low vision. Pre-orders start in the United States and South Korea, require an 18+ Google account from those countries, and work with specified Pixel, Samsung or iPhone models, with retail availability on Dec 10. Google said a capture indicator light is required, bystander faces are designed to be redacted before Gemini processes visuals, and those visuals are not used to train models or target ads.
AWS open-sources AWS Emulator, covering 200-plus services and 400-plus API models, written and tested by agents, free to run offline; an AWS executive said a replication engine that once took four months was rebuilt in about two hours once tests existed
Boss lens: letting coding agents trial AWS calls locally may cut cloud test spend. AWS itself says this is not a security test, so production still needs a pass on real AWS (inference).
CollapseAWS open-sources AWS Emulator, covering 200-plus services and 400-plus API models, written and tested by agents, free to run offline; an AWS executive said a replication engine that once took four months was rebuilt in about two hours once tests existed
AWS announced AWS Emulator on Nov 11 (the official blog added repository links on Nov 12): a local emulator of AWS APIs that the CLI, SDKs and coding agents can call on a laptop, with no extra service fee and source released under CC0. The AWS News Blog and a column by Swaminathan Sivasubramanian say it covers 200-plus services and 400-plus API models, was written and tested entirely by agents, and is gated by fidelity tests checked against real AWS; unimplemented operations return HTTP 501. Sivasubramanian also wrote that when he used a coding agent to rebuild a replication engine from an early distributed lock service, progress stalled for four nights until the agent had a spec and a test environment, after which the rebuild took about two hours—versus about four months the first time. AWS warns that the emulator does not evaluate IAM policies the way production AWS does, should not be used as a security test, and that applications still need to be tested on real AWS before launch. The announcement is dated Nov 11, just outside this edition’s main window, and is included as a 48-hour catch-up.
MiniMax open-weights H3-Image, saying it can generate a 4K image in under 10 seconds with up to 16 reference images, halves 4K-and-above requests in launch week, and prices at about 50–70% of mainstream peers
Boss lens: ecommerce and design teams can trial 4K quality during the half-price week. If you want to ship the weights inside a customer-facing product, have legal read whether the license covers your markets and commercial use (inference).
CollapseMiniMax open-weights H3-Image, saying it can generate a 4K image in under 10 seconds with up to 16 reference images, halves 4K-and-above requests in launch week, and prices at about 50–70% of mainstream peers
MiniMax dated its official blog post on H3-Image to Nov 11, 2026, and released the weights. The company said the model supports text-to-image, reference-based generation and editing, can produce a 4096×4096 (4K, about 16 megapixel) image in under 10 seconds, accepts up to 16 reference images, and can be up to 5x faster than mainstream alternatives. Weights are on Hugging Face, and the model is also offered through the MiniMax API, MiniMax Code and Hailuo AI. Official pricing language says Standard mode is 1.5–2x as fast as mainstream options and Highspeed 3–4x, at about 50–70% of comparable prices, with a launch-week 50% discount on 4K-and-above requests. License terms reportedly restrict some use in the US, EU, UK and South Korea, and commercial or hosted use may need MiniMax authorization; read the license itself. The post is dated Nov 11 and is included as a 48-hour catch-up.
Thinking Machines releases Inkling 2 Small weights under Apache 2.0: 594 billion total parameters, 17 billion active, available to download, fine-tune and self-host, with integrations including vLLM
Boss lens: if compliance requires models to run in your own environment, have engineering add this Apache 2.0 checkpoint to the shortlist. It is not a finished chatbot and will need people to deploy and maintain it (inference).
CollapseThinking Machines releases Inkling 2 Small weights under Apache 2.0: 594 billion total parameters, 17 billion active, available to download, fine-tune and self-host, with integrations including vLLM
Thinking Machines Lab updated its official blog on Nov 12 to say Inkling 2 Small’s weights are now on Hugging Face under Apache 2.0, so they can be downloaded, fine-tuned and self-hosted. The model is a mixture of experts with 594 billion total parameters and 17 billion active, and supports image input plus variable reasoning effort. The hosted model was introduced on Nov 5, when the lab said weights would follow after external safety testing; this update fulfills that, and says the model is integrated with Unsloth, SGLang, vLLM and TokenSpeed, and available via API on NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare. The lab cautions that it has no audio, is not optimized for chat, and can pick the wrong tools in skill-heavy codebases. For teams that want local deployment or fine-tuning, it is a new open-weight option, though still a technical evaluation.
Reuters: Chinese AI firms are rushing to rent Nvidia compute at overseas data centers, fearing the US may close a remote-access gap, according to anonymous sourcesUnconfirmed
Boss lens: if your training or inference depends on renting overseas Nvidia capacity from China, have procurement map contract terms and alternatives. The Senate has not passed the bill, so this is not an enacted ban, but renewals are a moment to read the fine print (inference).
CollapseReuters: Chinese AI firms are rushing to rent Nvidia compute at overseas data centers, fearing the US may close a remote-access gap, according to anonymous sources
Reuters reported on Nov 12, citing three people familiar with the matter, that some Chinese AI companies are trying to rent Nvidia-powered capacity at data centers outside China, fearing the controversy around Nvidia could push Washington to tighten rules. The report notes that while the US already restricts exports of the most advanced AI chips to China, Chinese firms can still rent those processors remotely at overseas facilities. The people said a major Chinese cloud provider has discussed locking in overseas capacity, some firms have contacted data-center operators about reservations, and smaller labs are considering compute providers in Southeast Asia. They pointed to the Remote Access Security Act, which passed the US House in January 369-22 and awaits Senate action, and would let the Commerce Department regulate remote cloud access to restricted technology. Reuters said the people declined to be identified or to name the cloud provider because of the sensitivity. This item was checked against a Reuters-credited reprint in The Hindu, not the paywalled Reuters original, and the core claims rest on anonymous sources, so it is marked unverified.
Four US nonprofits sue OpenAI, Anthropic, Google, Meta and SpaceXSI in San Francisco, asking a court to freeze further development of more advanced AI models
Boss lens: this is a filed lawsuit, not an injunction already in force, so it is not a reason to shut off tools you already use. If you depend heavily on the newest models from one of the five defendants, have counsel watch for any temporary limits or contract-language changes (inference).
CollapseFour US nonprofits sue OpenAI, Anthropic, Google, Meta and SpaceXSI in San Francisco, asking a court to freeze further development of more advanced AI models
On Nov 12, Irreplaceable Education Fund, Public Citizen, Heat Initiative and The Alliance for Secure AI filed a lawsuit in San Francisco Superior Court against Anthropic, Google, Meta, OpenAI and SpaceX’s AI unit SpaceXSI. In a Public Citizen release, the plaintiffs said they are invoking California’s Unfair Competition Law and asking the court to freeze further development and deployment of more advanced AI systems that, they argue, could escape human control and heighten public-safety risks such as bioweapons and large-scale cyberattacks. The release said the complaint cites cases in which existing systems bypassed safeguards, accessed outside systems without authorization, or took harmful actions, and points to public warnings by some industry figures about catastrophic risk. Former New Jersey Attorney General Matthew Platkin filed the complaint for the plaintiffs. The case has been filed; a court has not ruled on an injunction, and a filing is not a finding of liability.