SecondSourceJudgment rebuilt from primary sources

Alibaba launches Qwen4-Max; cost figures are self-reported

TODAY

Only the Tip Shows

  1. WEIGHTS of Alibaba's new flagship, kept in-house for now
  2. GUIDANCE NVIDIA's outlook, which an investor says may not come
  3. 3 of 70 tasks Surge AI made public in its spreadsheet test

At a glance

1. Alibaba releases Qwen4-Max without opening the flagship model's weights.

2. CNBC reports Treasury Secretary Scott Bessent saying NVIDIA should pick an outsider as CEO.

3. Investor Gene Munster says NVIDIA's quarterly report may skip guidance.

Correction: main-line item 1 of the November 14 issue covered ChatGPT Wallet, OpenAI's payment feature, and one sentence near its end rested on a wrong premise. That sentence pictured United Airlines reversing course by year-end to let AI agents book tickets, while fraud checks still often declined Wallet purchases. In that case, it said, payments would remain what holds agents back. The United terms quoted by Skift, a travel-industry outlet, ban unauthorized third-party personal AI agents, not all agents, so there is no blanket ban to reverse. The item's main judgment stands. The archive page has rewritten that sentence and added the scope of the ban to the paragraph on United's terms (November 14 issue).

Today's main line

1. Alibaba releases Qwen4-Max without opening the flagship model's weights

Concept illustration: two flat-drawn storefronts stand side by side on a cream background with a taller pale pink block behind them; the larger left shop has a sign lettered ALIBABA with an orange curved mark to the left of the word, and on its dark table sit a laptop whose black screen shows a white speech bubble, tagged QWEN4-MAX, and a thick red-orange book tagged WEIGHTS, with a tall white door-like panel at the right of the room. The smaller right shop has a sign lettered HUGGING FACE with a yellow smiling face with open hands to the left of the words, and on its light counter a smaller laptop showing the same kind of speech bubble, tagged QWEN3.8-MAX, and a dark grey-brown book tagged WEIGHTS, each tag joined to its object by a thin line. Printed on the image: Qwen3.8-Max weights are on Hugging Face; Alibaba previews open weights only for smaller Qwen4 models.
Alibaba released the full version of flagship model Qwen4-Max on November 16 and so far offers Qwen4-Max only on QwenCloud, Alibaba's platform for serving Qwen models, through the API and a Token Plan subscription. Alibaba self-reports that average cost per task on Work360, a work-task benchmark Alibaba built for this release, falls from US$4.43 for the previous flagship, Qwen3.8-Max, to US$1.67 for Qwen4-Max. The drop is spending per task on the benchmark, not a price-list reduction, and no other organization has rerun Work360.

In the full story: how the price compares with the previous flagship, how the cost cut was measured, and what was open-sourced this time

2. CNBC reports Treasury Secretary Scott Bessent saying NVIDIA should pick an outsider as CEO

Concept illustration in two side-by-side panels: on the left, in front of a columned building with a sign lettered US TREASURY, a person looks through a video camera on a tripod aimed at a woman in a dark coat with a lanyard badge, who stands beside a microphone on a stand. On the right, under a sign with a green square emblem and the lettering NVIDIA, an empty red office chair lettered CEO stands at the end of a long table where two people sit reading open papers. Printed on the image: NVIDIA's board picks the CEO; the Treasury Secretary's view is only an opinion.
According to CNBC, US Treasury Secretary Scott Bessent said in a Fox News interview on November 15 that NVIDIA should pick someone from outside the company to succeed Jensen Huang, whose resignation as CEO was accepted on November 12. The same day, a Financial Times report, relayed by Fox Business and CNBC, said Huang has discussed a return with close associates and that an unnamed ally described a plan to seek White House support. Seeking White House support is still only the ally's plan, and so far only the Financial Times has reported the comeback push.

Our inference: since Jensen Huang resigned, the choice of his successor has pulled in Washington, and the board must weigh Washington's reaction before choosing.

In the full story: what Bessent actually said, what Huang's camp is doing, and how the board responded

3. Investor Gene Munster says NVIDIA's quarterly report may skip guidance

Concept illustration: a framed white display board stands on a black base, with a black band across the top bearing a person's name in large white capital letters. Inside, ten columns of bundled paper stacks, four high and each tied with a white band, rest on a lower black band that has a green square emblem and the lettering NVIDIA; the nine left columns are beige and the rightmost column is red. Printed on the image: About 10% Munster's estimate: share of NVIDIA revenue at risk if US Commerce tightens limits on Asia sales.
Investor Gene Munster wrote on X on November 15 that NVIDIA may give no next-quarter guidance when NVIDIA reports quarterly results on November 17. One reason Munster gives is that the management change makes owning the guidance hard for a new CFO, whom Munster does not name. Munster's estimate of the revenue share affected is not a regional share NVIDIA has disclosed, and Munster gave no method or time period for the estimate.

In the full story: the two reasons Munster gives, how much revenue faces controls, and what guidance NVIDIA gave last time

Named commentary

1. Researcher Lucas Beyer says claims that AI will take over humanity once it becomes superintelligent are science-fiction fan fiction.

Concept illustration on a salmon background: at left a book propped on a stand is lettered FANFIC, its cover showing a large laptop with a speech bubble above a cluster of small houses with small laptops; in the centre a simple round-headed figure sits at a long white table typing on a laptop whose screen shows a speech bubble, beside an empty chair facing a second laptop with a blank screen, and red cables from the two laptops join and plug into a round socket on tall server racks at the right. Large black quotation marks sit at the top, and a dark silhouetted person in the lower-left foreground points a finger toward the table. Printed on the image: Per Beyer, models produce tokens only when called and need large clusters, so no fast self-copying.
Researcher Lucas Beyer wrote on X on November 15 that claims that AI will take over humanity after becoming superintelligent are science-fiction fan fiction, a personal judgment with no data attached. Beyer argues that near-term risks such as job disruption and cyberattacks should come first, and that too is Beyer's own judgment.

Research & models

1. According to the paper's abstract, Surge AI's new GDP.xlsx benchmark finds the best of 20 model configurations met every grading criterion on one attempt in only 38.3% of 70 real workplace spreadsheet tasks.

Concept illustration: a white cabinet with two drawers has a framed display compartment on top, headed by a black banner with a purple-and-orange S-shaped mark and the lettering SURGE AI on the left and GDP.XLSX on the right. On the shelf below, from left to right, stand a black laptop whose screen reads GEMINI 4 ARGON, a small white table card with a multicoloured G and the lettering GOOGLE, a tall stack of beige sheets with a grid on top and a black bug sitting on it, and a lower stack of brick-red sheets with a grid on top. Printed on the image: 38.3% Spreadsheet tasks fully passed first try: top model vs 19.3% median of 20 configurations (Surge AI).
According to the paper's abstract, Surge AI, a company whose business is evaluating AI, tested 20 model configurations on a new benchmark called GDP.xlsx, made up of 70 real workplace spreadsheet tasks, and the top score went to Google's Gemini 4 Argon. Surge AI wrote the tasks and AI judges graded the work against rubrics written by experts, and the abstract gives neither the agreement rate between the AI judges and human experts nor how many runs each task got.

2. Operations-research scholars report in their paper's abstract that a scheduling solver written by an AI coding agent came closer to the theoretical lower bound than commercial solvers on a set of factory scheduling problems with no known optimal solution.

Concept illustration: on a large white panel resting on a dark brown base, a black computer tower and a monitor showing grey text lines sit at the left, beside two horizontal bars with white-on-black labels, "CODING AGENT" on a red bar and "CP OPTIMIZER" on a longer beige bar with a black outline. A vertical black line runs down through both bars beneath a white-on-black label reading "LOWER BOUND"; the red bar ends a short way past the line, while the beige bar continues almost to the panel's right edge. Printed on the image: 1.8% Agent-written solver's max gap over lower bound, vs CP Optimizer's 2.6% to 5.8%; authors' figures.
According to the paper's abstract, four operations-research scholars had an AI coding agent write a solver for factory scheduling problems with no known optimal solution, then ran the agent-written solver against the commercial solvers CP Optimizer and Hexaly on the same workstation with the same compute budget, measuring how close each came to the theoretical lower bound, below which completion time can't go. The scores are the authors' own and no one else has rerun the comparison, and the abstract doesn't say which language model was used, how much time and how many tokens the agent spent, or whether the commercial solvers were tuned.

3. A group of authors reports in their paper's abstract that adding process control outside the model, without changing the underlying model, raised the unattended completion rate on a 120-case benchmark the authors built from their own business workflows from 50% to 100%.

Concept illustration in a rounded frame topped by a black bar lettered ARXIV in white: on the left are two laptops, one above the other, each screen showing a grey speech bubble with grey lines, and each is joined by a long horizontal beige bar to a dark grey-brown tray of stacked beige folders on the right. The upper bar is plain, with a loose, untidy pile of beige folders below its middle, while the lower bar is lined along its whole length with a red picket railing and leads to a taller stack of folders. Printed on the image: 50%→100% Authors report: unattended completion on 120 cases, same model, before/after outside-model control.
A group of authors self-reports in a paper abstract adding process control outside an unchanged underlying model: workflow definition, state and domain knowledge move into versioned files, and a state machine with machine-checkable acceptance criteria guards each step at run time. The tests cover only the authors' own deployed workflows, each longer than 200 steps, and the abstract doesn't say what business the workflows serve, what baseline setup ran before the change, or whether the test cases overlap with the cases used to tune the workflow.

Product moves

1. Varun Mohan, a member of the team behind Antigravity, Google's coding-agent product, says the context windows of the three models in Antigravity (Argon, Carbon and Hydrogen) have doubled.

Concept illustration: a laptop screen is split into two halves, each under a red label reading CONTEXT WINDOW, with a dark grey arrow pointing from left to right between them; on the left a narrow red-outlined box is packed with seven grey-lined speech-bubble icons, a new bubble arrows in from the right, and an arrow out of the box's left side leads to a crumpled grey-white lump whose fragments fall onto a large pile of crumpled pieces, while on the right a wider box holds only three spaced-out bubbles and the same flow ends in a much smaller pile. Above the laptop at top right is a dark panel with a red, yellow, green and blue letter G and the lettering ANTIGRAVITY GOOGLE, and two hands rest on the keyboard at the bottom. Printed on the image: 1/3.5 Mohan self-reports: Antigravity compactions per user turn after windows doubled, vs before.
Varun Mohan, a member of the team behind Antigravity, Google's coding-agent product, says in a post that the context windows of Argon, Carbon and Hydrogen, the three models in Antigravity, have doubled. The compaction count per user turn is self-reported by Mohan, and the post gives no measurement method, no window sizes before or after the doubling and no long-task success rate.

2. Apple is starting a US research study on AI for health, enrolling up to 200,000 people.

It is built on technology from OpenEvidence, which Apple calls a clinical AI platform, and it is a study, not a product launch (Apple Newsroom, 2026-11-15).

Also happened

1. Not verified by us yet: Trade reporter Ana Swanson, relaying a New York Times report on X, says US Commerce Secretary Howard Lutnick pushed to loosen a chip rule affecting data centers in Southeast Asia, a rule that overlaps with the Lutnick family's business (Ana Swanson, 2026-11-15).

2. Not verified by us yet: Bloomberg reporter Mackenzie Hawkins, relaying a Bloomberg report on X, says Taiwanese prosecutors have indicted an NVIDIA manager in a case involving chip smuggling (Mackenzie Hawkins, 2026-11-15).

Sections with nothing new this issue

Chips & semiconductors: No chips & semiconductors item this issue. The two NVIDIA items on the CEO change and the upcoming earnings report are main-line items 2 and 3.

From the archive

SecondSource isn't a news digest: AI reads primary sources across fields, every judgment comes with its original source, and mistakes are corrected in public. The focus is what has changed and which claims still lack evidence, not what happened today.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.