Skip to content

← X AI Daily

X AI Daily

Date12 November 2026 · ThursdayWindow2026-11-11 02:00 to 2026-11-12 07:55CoverageComplete

Translation

English translation of the Chinese original.

1. AI Models, Products, and Applications

Qwen4-Max launches as a two-week preview: 1M context, Token Plan only, not open source

At 10:01 Beijing time on November 11, Alibaba's official Qwen account announced that Qwen4-Max has entered preview, with wording centered on "deeper knowledge, richer world understanding, and stronger coding and agent capabilities," aimed at complex programming, knowledge work, front-end building, and professional software workflows. The same day, Alibaba Cloud Model Studio put up a documentation page for qwen4-max-preview: 1M token context, maximum output of 393,216 tokens shared between thinking and the final answer under one budget, text, image, and video input, thinking on by default with reasoning_effort adjustable across low, medium, and xhigh, support for function calling, web search, and structured output, and an interface compatible with OpenAI's Chat Completions and Responses as well as Anthropic's Messages. On channels, the documentation states that during the preview it is offered only through Token Plan with no pay-as-you-go support, and that dedicated endpoints use Token Plan's dedicated domain; the preview runs from 10:00 on November 11 to 24:00 on November 25 (UTC+8), during which the model will not change and after which it goes offline. The documentation published the same day contains no weights, no model card, no benchmark table, and no parameter count.

Why it matters: this is the most widely spread AI post in this window, with 1.113 million views. But its shape is "a two-week window plus a subscription model plus no weights": developers can use it, yet they cannot get reproducible evaluations or a path to local deployment. The line in the documentation that the model retires when the preview ends was ignored in most retellings, and it determines whether anything built on it in these two weeks can keep running long term.

  • Core post: Alibaba_Qwen announcement post, about 1.113 million views, 11,754 likes, 1,458 reposts; AmControo's hands-on Qoder test, about 33,000 views, 226 likes, 114 reposts, saying they tried full-site generation in Qoder.
  • Verified: the Alibaba Cloud Model Studio model page confirms the 1M context, the 393,216-token output ceiling, multimodal input of text plus images plus video, Token Plan exclusivity and no pay-as-you-go support, the preview period from 10:00 on November 11 to 24:00 on November 25 with retirement at expiry, the thinking-effort parameter, tool calling and structured output, and compatibility with OpenAI and Anthropic API formats.
  • Source claims: that Qwen Chat, QwenWork, Qwen App, and QwenStudio also have entry points appears only in secondary reporting and product account posts; at the official documentation level, this edition verified only Token Plan. There is no official information on parameter scale, benchmark results, license, or open-source timeline.
  • Assessment: the information value of this launch lies mainly in how it is accessed rather than in capability — the company gave no comparable evaluation, the 1M context and 393K output are hard specifications, but "stronger" is for now only its own wording; the fact that the model goes offline in two weeks means product validation built on it now needs an exit path in place before November 25.

Google releases Gemini Decision 1: up to 1,000 questions answered per forward pass, output not billed

At 02:26 Beijing time on November 12, Google announced the Gemini Decisions API on Google AI Studio, built on a new model, Gemini Decision 1. It does not generate text token by token; instead it sends a shared input together with up to 1,000 closed-ended questions through a single forward pass and directly returns option probabilities and confidence scores from 0 to 1. Input can be text, images, audio, video, and PDF, up to 256K tokens; there are three question types — yes/no (optionally with true and false criteria), choose one of 2 to 255 (returning the winner and normalized probabilities for all candidates), and rating on an ordered scale (returning a weighted score and the probability of each band). Pricing is $0.10 per million input tokens with output tokens not billed at all; the shared input is counted once per request, and questions only add the tokens of their own instructions and criteria. Official end-to-end latency is 150 to 250 milliseconds, of which model compute is a median of 50 to 100 milliseconds, essentially flat as the number of questions in a request rises from 1 to 100 within a 16K context; on the same shared benchmark it is nearly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API.

Why it matters: this is a new product form that pulls "judgment" out of generative calls. The highest-volume links in agent systems — tool selection, policy checks, context triage, model routing — are exactly the ones that used to be most expensive and slowest, and those calls never needed a natural-language answer. The pricing structure of not billing output directly changes the cost model for such systems.

  • Core post: Google announcement post, about 225,000 views, 1,231 likes, 103 reposts; Artificial Analysis evaluation post, about 39,000 views, 412 likes, 20 reposts.
  • Verified: Google's official blog (November 11, 2026) confirms the 1,000-question cap, 256K input, the three question types and what each returns, $0.10 per million input tokens with zero output billing, 150 to 250 milliseconds end to end and 50 to 100 milliseconds of model compute, latency staying essentially flat as question count rises, and being nearly five times cheaper than Flash-Lite plus the Interactions API; VentureBeat's report confirms it ranks 9th among 18 decision models on the Artificial Analysis Decision Index v2.1.
  • Source claims: the specific score of 66, the cost of $345 per million tasks, and trailing Drex Max v1.1 (73 points, $103) and GLiDE v1.2 (73 points) by 7 points all come from Artificial Analysis's own post; that organization's corresponding article page would not open in this edition.
  • Assessment: the real change is in the interface form and the pricing method, not in the leaderboard position — 66 points is only mid-pack, and it trails on Artificial Analysis's own classification and extraction benchmark AA-Paperclip (33% of documents fully correct versus 63% for Drex Max), indicating that "fewer tokens, lower latency" comes at the cost of complex document handling; treating it as a router and gatekeeper is reasonable, treating it as the workhorse for document understanding is not yet established.

OpenAI publishes CoT monitorability research: the stronger the model, the less the chain of thought can serve as a monitoring tool

The OpenAI Alignment site posted the research "What affects chain-of-thought monitorability across the model lifecycle?" on November 10, with authors including Jeffrey G. Wang, Aiden Low, Zora Che, Holden Mui, and Harshit Sikchi, and corresponding authors including Micah Carroll, Ben Sokolowsky, Katherine Lee, and Jakub Pachocki. One of the authors, Jeffrey Wang, introduced it in a post on X at 02:17 Beijing time on November 12, saying the conclusion is "unsurprising but sobering." The research starts from three observations on GPT-6 Astra: CoT monitorability declines under non-adversarial conditions, the ability to evade CoT monitoring grows in adversarial testing, and no-CoT capability is markedly higher than in earlier models (the paper cites external evaluation by the UK AISI). The research treats no-CoT capability (the ability to complete a task in a single forward pass without emitting any chain-of-thought or tool-call tokens) and CoT controllability (the ability to add to, remove from, and manipulate its own chain of thought on request) as proxies for worst-case monitoring-evasion capability, and breaks this down into four stages along the training lifecycle: in pretraining, no-CoT capability is a smooth function of base-model perplexity, so better models are more capable, while low perplexity and higher CoT controllability are only weakly correlated; in architecture, most no-CoT evaluations are explained by perplexity and are architecture-independent, but on multi-hop planning and serial-computation evaluations some architectures are clearly better at the same perplexity, with steeper curves; in reinforcement learning, CoT controllability rises even with no process supervision beyond length constraints and final-answer scoring, and rises more when tool access and tool supervision are added; finally, when fine-tuning on model-generated data, CoT controllability can be passed down to the next-generation model along the training data, much like latent learning.

Why it matters: CoT monitoring is a supervision method that is often relied upon today. This research amounts to acknowledging that as models get stronger, this pillar naturally decays, and that part of the decay comes from training choices they actively made — RL and distillation both push in the same direction. For teams deploying agents, this means "reading the chain of thought" cannot serve alone as the pre-launch safety gate.

  • Core post: Jeffrey Wang's post, about 101,000 views, 879 likes, 126 reposts, 502 bookmarks.
  • Verified: the original OpenAI Alignment text confirms the publication date, the author list, the definitions of no-CoT and CoT controllability, the specific direction of the four conclusions (including "controllability rises during RL" and "controllability can be transferred through SFT"), the motivation of three regressions on GPT-6 Astra, and the authors' own line that "models are becoming harder to monitor through their CoT"; the definition of the CoT controllability evaluation suite and its original conclusions appear in Chen et al.'s paper.
  • A framing difference that needs flagging: OpenAI says controllability rises during RL, while the March paper it cites measured the opposite direction on open-source models (controllability falls as RL training and reasoning length increase). The two measured different models and pipelines; the body text follows this release's account but should not be treated as an independently replicated conclusion.
  • Assessment: this is an internally honest piece of research that is also a risk notice — it offers no alternative, three of its four findings point toward "continuing to scale on the current paradigm will make monitoring harder," and the strong correlation between no-CoT capability and perplexity shows precisely that this capability cannot be stripped out by changing architecture alone.

DeepMind releases AlphaGenome 2: from tissue averages to single cells, trained on 107 million cell profiles

Google DeepMind released AlphaGenome 2 on November 11, changing the previous generation's approach — "predict thousands of experiments in parallel from one stretch of a million bases of DNA" — into a conditional model: in addition to the DNA sequence, tissue, cell type, or cell state is fed as a conditioning input into every layer of the architecture, and it predicts the outcome under only one condition at a time. Training data is human and mouse bulk and single-cell data; the official text says more than 107 million single-cell profiles, while the introduction says "over 100 million cell profiles spanning 68 tissues." In capability, it predicts gene expression, histone modifications, and chromatin accessibility, and adds DNA methylation; it can do cross-modal imputation, inferring unmeasured chromatin accessibility from a single measurement (such as single-cell gene activity); and it can condition on unseen cell states without retraining, with scientists simply uploading single-cell snapshots of a few dozen cells. The official results are that on TenK10K (a single-cell atlas of 28 primary immune cell types) causal variants identified rise from 16% to 30% and directional gene effects from 23% to 43%, and that it reproduced the existing finding that the ETS2 distal regulatory variant in inflammatory bowel disease is over-activated only in intestinal macrophages. On availability: from November 11 it is open to non-commercial research through the AlphaGenome API and Colab tutorials, commercial research goes through a private preview, and AlphaGenome 2 Skills will later be offered in Google Antigravity. No weights were released.

Why it matters: genome-wide association studies can find disease-associated stretches of DNA, but often cannot find which specific variant it is, which gene it acts on, and in which cell type it acts, and this gap has long been the main obstacle to connecting genetics with cell biology. Shrinking the unit of prediction from "tissue average" to "cell type and cell state" is a rare structural improvement in this chain rather than a small step up in scores.

  • Core post: Pushmeet Kohli's post, about 34,000 views, 1,153 likes, 130 reposts. Published at 23:25 Beijing time on November 11.
  • Verified: the DeepMind official blog confirms the conditional-model design, the training scale of 107 million single-cell profiles and 68 tissues, the added DNA methylation prediction, cross-modal imputation and zero-shot conditioning on unseen cell states, the improvements on TenK10K of 30% versus 16% and 43% versus 23%, the reproduction of ETS2 and inflammatory bowel disease, and the availability model of API for non-commercial research and private preview for commercial use; the accompanying paper PDF is the source for details, and The Verge's report provides a third-party description.
  • Source claims: the blog's introduction and body use two different formulations for training scale ("over 100 million" and "107 million"), and the company did not further break down how cell profiles and tissues are counted.
  • Assessment: this is a textbook step in moving foundation-model methodology into genomics, and the conditional design lets the model predict unseen cell states, which is worth more than a few more benchmark points; but it is API plus private preview without weights, so reproducibility depends on DeepMind itself, and the recall gains for "causal variants" are still readings on a specific single-cell atlas and should not be extrapolated to all cell types.

Runway open-sources Praxis-1 under Apache 2.0: a 27B world action model, 46.2% on a simulation benchmark

On November 11, Runway moved Praxis-1 from early-partner testing to public release, with weights, inference code, and full fine-tuning and LoRA fine-tuning code all released under Apache 2.0, commercial use included. The model is a 27B-parameter mixture-of-experts diffusion Transformer that predicts video and robot actions at the same time, on the idea of turning what large-scale video pretraining learns about "how objects move and what a task looks like halfway through" into control signals. The company says it runs out of the box on data-center cards such as the H100 and B200, on workstation cards such as the RTX 5090 and RTX PRO 6000, and on Jetson Thor mounted on robots; download channels are GitHub, Hugging Face (with LeRobot support), and vLLM-Omni, with hosted inference also available. On results, the company says it is the highest-success-rate world action model on NVIDIA's RoboLab-120 at 46.2% (1,200 episodes), versus 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano, and 28.0% for π0.5, and that it outperforms all open-weight vision-language-action models; another set of numbers is 80.0% success on SimplerEnv after fine-tuning on a Google RT-1 robot that never appeared in the training data, versus 52.4% for the original RT-1 policy. Early partners in this release include Noble Machines, Standard Bots, and Ultra, running on a humanoid, a six-axis industrial arm, and a dual-arm warehouse robot respectively.

Why it matters: robotics has long lacked data rather than model architectures, and Runway's bet is that "internet video can substitute for teleoperation video in pretraining," which this week has a downloadable implementation. More important is the license: Apache 2.0 including commercial use means hardware makers can fine-tune directly on their own bodies without handing over their data.

  • Core post: Runway announcement post, about 47,000 views, 417 likes, 76 reposts; benchmark explainer post, about 6,082 views, 41 likes. Both published at 23:04 Beijing time on November 11.
  • Verified: the Runway official release page confirms the 27B mixture-of-experts diffusion Transformer, Apache 2.0 covering weights plus inference and fine-tuning code and including commercial use, 46.2% on RoboLab-120 and 42.9%, 36.8%, and 28.0% for the three comparison models, 80.0% versus 52.4% on SimplerEnv, the three channels of GitHub, Hugging Face, and vLLM-Omni, and the hardware range from H100 to Jetson Thor. RoboLab-120 itself is NVIDIA's simulation benchmark (code repository, project page and leaderboard), with 120 tasks and multiple trials per task, so 46.2% is a simulation result rather than a real-robot result.
  • Source claims: the 46.2% comes from the scores Runway published, and this edition did not verify a corresponding entry on NVIDIA's leaderboard page, nor is there independent replication. Note also that Runway's September introductory page said at the time "public release in a few months" (introductory page); the November 11 release is the one that actually shipped the weights.
  • Assessment: open weights plus a commercial license is the most practically useful thing for hardware makers this week; but the benchmark is a simulation benchmark and the results are vendor-reported, so 46.2% should be read as "leading under its own evaluation setup" rather than "better than all open VLAs on real robots."

2. AI Compute and Industry

Altman tells Samsung executives AI demand is being "double counted," while remaining bullish long term

South Korea's Financial News ran an exclusive report at 17:30 on November 11 saying it had obtained Q&A material from a meeting between OpenAI CEO Sam Altman and Samsung executives: the meeting was on October 31 at the Samsung Human Resources Development Center (Yongin), with about 50 presidents of Samsung affiliates in electronics, finance, and biotech attending. Asked whether demand for AI, memory, and data centers would saturate or slow over the next 5 to 10 years, he answered that "there will be short-term fluctuations"; referring to this year's bubble debate around neoclouds (emerging cloud providers that rent out AI compute), he said "there is double counting in demand," and that "the only two companies really driving large demand growth are OpenAI and Anthropic, so there could be a supply shock and a period of turbulence." Long term he remains bullish, saying AI demand will grow substantially, and gave the example that in early 2020 the heaviest token users used about 100,000 tokens a month, while at the start of this year the heaviest users used 100 billion tokens a month. Asked about the biggest bottlenecks to expansion, he listed the chip supply chain, financing, AI safety and social acceptance, and enterprise transformation, and stressed that the biggest constraints in the coming years are the chip supply chain and data-center infrastructure, which need more investment.

Why it matters: the person saying this is pushing the largest-scale compute expansion, and his audience is the world's largest memory supplier, while "double counting" points at a very concrete problem — the same data center or the same batch of compute often gets counted repeatedly under multiple projects' accounting. The talk of a supply shock, placed alongside his own call for more investment, is the real information in this report: the short-term demand curve is unreliable, the long-term direction is unchanged.

  • Core post: Wall St Engine's retelling, about 283,000 views, 729 likes, 69 reposts, published at 21:16 Beijing time on November 11.
  • Verified: the original Financial News article confirms the meeting time and place, the attendance of about 50 affiliate presidents, the two quotes "there is double counting in demand" and "only OpenAI and Anthropic are really driving growth," the language about a supply shock and a period of turbulence, the two comparison figures for token usage, and his judgment on the chip supply chain and data-center infrastructure.
  • Source claims: the entire report is based on Q&A material the outlet obtained, and neither OpenAI nor Samsung has publicly confirmed it; this edition also did not find a second outlet that obtained the same material.
  • Assessment: this reads more like expectation management aimed at customers and suppliers at once — asking Samsung to keep investing in capacity while warning it not to take industry-wide demand forecasts at face value; reading it as "Altman is bearish on AI" is overinterpretation, since both his long-term judgment and his call for investment are explicit in the report.

Follow-up: Anthropic roadshow material says Claude Mythos got Claude running on AMD MI355X in about 3 days

We already reported yesterday that Anthropic has begun its IPO roadshow; here we add one widely retold technical disclosure from the roadshow video. In the video, chief compute officer Tom Brown says Anthropic now trains and serves Claude across four classes of chips, and the slide lists AWS Trainium, Google TPU, NVIDIA GPU, and AMD GPU; in the past, getting a new chip platform running took "dozens of engineers months of work," and this time it was handed to Claude Mythos to try, and "it took about 3 days to get Claude running on AMD, with basically no human intervention," adding that "Claude ported or wrote nearly 100% of the kernels." A slide footnote narrows those three days considerably: it refers to the approximate time needed for Claude Mythos Preview to bring prompt processing for one particular Anthropic model on AMD MI355X up to a usable performance level; the 6-to-12-month comparison is the company's historical experience bringing new accelerators online with conventional engineering, and the footnote does not cover training and gives no speed figures. The video also states that Mythos is deployed only to cyber defenders and is not a public model.

Why it matters: if the software stack for new hardware can be compressed to a few days in the hands of this class of model, then the cost structure of "locking into one chip vendor" loosens, and both procurement leverage and the feasibility of multi-chip parallel inference change. This is exactly the part that got amplified in this week's retellings into "the CUDA moat is gone," but the video itself does not say that, and the narration does not mention NVIDIA or CUDA.

  • Core post: MilkRoadStocks's retelling, about 1,261 views, 14 likes, published at 03:15 Beijing time on November 12; getsuperintel's qualified retelling, 395 views, 5 likes, explicitly noting that the footnote covers only prompt processing for a single model and contains no speed figures.
  • Verified: CellCog's word-by-word check of the roadshow video's subtitles and slides confirms the video length of 42 minutes 43 seconds, its location on Anthropic's investor site, Tom Brown's two quotes and the footnote text, the list of four chip classes, and that Mythos is deployed only to defenders; The Decoder's report and OfficeChai's report give descriptions in the same direction.
  • No primary source obtained: fetching Anthropic's investor site failed in this edition, the roadshow video and the filing document were not read, and the quotes and footnote above come from third-party verbatim transcription and should be treated as source claims.
  • Assessment: this is a disclosure amplified by retelling — the actual information is "prompt processing for a single model reached usable performance on MI355X in about 3 days," not "Claude runs fully on AMD" or "the training migration is complete"; "the CUDA moat is gone" is the retellers' conclusion, not Anthropic's statement.

Ramp's October AI Index: enterprise AI spending up 1.65% month over month, engineering teams still 72%

Ramp published the October edition of its AI Index on November 11. The data shows that across the more than 85,000 US businesses using Ramp corporate cards and bill pay, AI spending rose 1.65% month over month in October, a monthly record; in token terms, OpenAI rose 2.1% month over month and Anthropic 0.5%. This edition adds a breakdown by department: engineering teams account for 72% of all categorized spending, still the largest category but down from its summer peak; customer success and support teams rose from 0.8% in June to 13.3%. On price, the average price per million tokens fell 21% in a month, and token volume keeps growing faster than token spending. On concentration, the top 1% of businesses account for about 80% of the revenue OpenAI and Anthropic get from Ramp customers, with a median spend per person of $10,200 a month for these businesses (up 9% in October), $801 for the top 10%, and $16.32 for the median business. Frontier models (Opus, Fable, Sol, Astra) account for 69.5% of spending, up from 67.7% last month.

Why it matters: this is one of the few indicators that measures AI demand from actual payment data, and what it now says is "volume is rising, money is not": the 21% drop in price per token has eaten the usage growth; Ara Kharazian's conclusion in the post is that the market should turn its attention to new directions in AI transactions.

  • Core post: Ara Kharazian's post, about 93,000 views, 292 likes, published at 00:38 Beijing time on November 12.
  • Verified: Ramp's official November AI Index confirms October spending up 1.65% month over month, token spending at OpenAI and Anthropic up 2.1% and 0.5% respectively, engineering teams at 72%, support up from 0.8% to 13.3%, average price per million tokens down 21%, frontier models at 69.5% share, $10,200 per person for the top 1% and $16.32 for the median business; the index explainer page describes the sample scope and the caveat that it "may understate adoption and spending."
  • Source claims: the post says "AI spending rose 2% in October," while the official figure is 1.65%; the statement that "OpenAI and Anthropic growth slowed to low single digits" is consistent with +2.1% and +0.5%, but both are still growing, not declining.
  • Assessment: this data matters more for the compute-investment narrative than for the model companies — with unit prices down 21% and spending up only 1.65%, usage must sustain an extremely high compound growth rate to support current infrastructure expectations, and the new share is currently coming from scenarios like customer service and support.

a16z proposes a "B2A economy": agents become the internet's main users

Olivia Moore and Justine Moore of a16z published "Welcome to the B2A Economy" on November 11, arguing that after business-to-business and business-to-consumer (B2B, B2C) a third class of customer has appeared — the agent. The piece opens with three claims: agents already account for the majority of web traffic, they write more than half of new articles published online, and they generate most of the code pushed to production. It then gives a set of traceable figures: HUMAN Security observed agent browser traffic growing 8x from August to October, with Muse taking the largest share at 60% and OpenAI products 30% combined; Shopify saw AI-driven store traffic up 8x year over year in Q1 2026, with orders from AI search up nearly 13x; in September 2026 more than 40% of PRs on GitHub began as agent drafts, versus 17% six months earlier; Stripe said in June that agents accounted for nearly 40% of API documentation traffic and 70% of CLI requests; Vercel said in October that agent-triggered deployments grew 28x in a year. The piece sees four categories of beneficiaries: agents and agent-native models, new products spawned by agents, incumbent platforms that optimize for agents, and the infrastructure agents need (payments, identity, communication), and predicts that agents will soon become the primary customer for most products.

Why it matters: this is a framework that treats "agent traffic" as a distribution channel rather than a feature. Its most direct practical implication for companies is counterintuitive: after years of retail and platforms building walls to block crawlers, they now have to compete for agents' business; the piece also points out an underrated current state — most agent traffic still goes through browser clicks, and most companies still block agent traffic out of fear of fraud and chargebacks.

  • Core post: a16z data post, about 69,000 views, 908 likes; Olivia Moore's main post, about 101,000 views, 346 likes; HUMAN Security's 8x data post, about 14,000 views, 96 likes; 0xJuliechen's discussion of the interface layer, 4,356 views, 36 likes, pointing out that how agents reach people is the other half of the story.
  • Verified: the full text appears on the a16z official site and the a16z.news version, confirming the authors, the four-beneficiary framework, HUMAN Security's 8x and Muse's 60%, OpenAI products at 30%, Shopify's 8x and 13x, GitHub's 40% versus 17%, the Stripe, Vercel, and Neon figures, and the caveat that "most companies still block agent traffic."
  • Source claims: the three statements "agents account for the majority of web traffic," "they write more than half of new articles," and "they generate most production code" are the article's core assertions, but the text gives no source for each item and the graphics carry no source labels; this edition did not find a public report supporting the original basis for these three.
  • Assessment: what can be verified is the trend and individual platforms' self-reported data (8x, 40%, 28x); what cannot be verified is aggregate judgments like "the majority" — none of the three assertions carries a source, and there is tension when they sit alongside "most companies still block agents"; reading it as a definitional piece for an investment theme is appropriate, citing it as statistical fact goes too far.

3. Chinese AI Ecosystem

ERNIE 5.5 Flash Preview released: 280B total parameters, 14B active, 1M context, weights said to be open-sourced before November 25

Baidu released ERNIE 5.5 Flash Preview on November 11, the first post-trained model built on ERNIE 5.5 Flash Base released on November 5, with 280B total parameters, 14B activated per inference, and 1M token context, targeting coding, 3D modeling, knowledge work, and content creation. The API is already live on the Qianfan platform, and the company says decoding speed remains fast. On licensing, the official page's wording is that "the weights will soon be open-sourced under Apache 2.0," with no date given. The official demos are all completed end to end in agent environments: making a third-person wuxia game, "Yunxiu Xing," in Unreal Engine 5.8 through a single continuous 11-hour coding task, merging cross-version art assets into one project, writing the full logic in C++, and turning the floating characters, inconsistent running direction, and misplaced camera discovered during playtesting into runnable code tests; generating nearly 100 mobile mini-games from a design document; building Daxing Airport along with a 747 and building Horyu-ji using Blender's Python API; making an editable W16 engine model in FreeCAD; and producing a sci-fi disaster film blockout previz with about 2,100 objects and 1.82 million triangles. Another topic in Chinese-language circles is anonymous testing of Qwen4-Max: Qoduck, which ran anonymously in Qoder for 4 days, was revealed on November 11 to be Qwen4-Max-Preview, and some users used that to infer the performance of the official release, arguing the preview falls short of expectations.

Why it matters: this is the most substantive release on the Chinese model side this week, and the point is not the parameter count but the workflow of "using code to patch defects it cannot see itself" — a pure text model discovering floating characters and misplaced cameras in a game engine by writing tests shows that the bottleneck for long-horizon agent tasks is shifting from generation capability to self-verification capability. Baidu's cadence is also worth noting: the base model on November 5, the post-trained preview on the 11th.

  • Core post: Baidu_Inc announcement post, about 179,000 views, 1,353 likes; Baidu_ERNIE parameter post, about 12,000 views, 255 likes; OpenRouter free-access post, 1,292 views, 41 likes; Chinese-language hands-on tests and retellings such as xiaohu, 7,708 views, 16 likes, vista8's breakdown of parameters and architecture, 4,555 views, 10 likes, canghe's impressions of "Yunxiu Xing", 4,755 views, 16 likes, JangXiin mentioning the November 25 open-sourcing, 2,262 views, 25 likes; the Qoduck discussion cites ai_funss's post, 4,108 views, 9 likes, saying Qoduck is the anonymous test of Qwen4-Max-Preview and that results are underwhelming.
  • Verified: the ERNIE official release page confirms the 280B total parameters, 14B active, 1M context, post-training on the November 5 Flash Base, availability on the Qianfan platform, weights to be open-sourced under Apache 2.0, and the specific process and external asset sources for the "Yunxiu Xing," nearly 100 mini-games, Blender, and FreeCAD demos; the base model release page confirms the 280B sparse MoE with 14B active architecture that this derives from.
  • Account-stated: the two items of open-sourcing the weights before November 25 and 200+ tokens/s per user on Qianfan come from the official Baidu_ERNIE account, and free access on OpenRouter until November 30 comes from another post on the same account; the official release page says "soon" and gives no speed figure, and the OpenRouter page was not verified in this edition; the license is stated only as Apache 2.0 with no text attached, while the November 5 base model was described in retellings as using a custom license excluding the EU and the UK — the two accounts are inconsistent, and the license text was not verified in this edition.
  • Assessment: the model's usability has to wait until the weights are actually out to judge well, but the quality of this release's demos is clearly above the usual "benchmark scores plus screenshots" — 11 hours of continuous coding and writing its own tests to fix bugs are claims an outside party can reproduce and check; the speed, the open-sourcing date, and the free window currently rest only on official account statements, and the release page does not state them.

This edition did not obtain Anthropic's original roadshow material; low-engagement releases such as n8n Nod, NVIDIA Alpamayo 2 Edge, and MBZUAI PAN 2 were not listed separately.