Model

Baidu's open weights carry a $50 million tripwire, and its investors are waiting for it to fire

Baidu released a 280-billion-parameter open-weights base model under a license that goes free only below $50 million a year in deployer revenue. That trigger, not the benchmark table, is the story, because it arrives while a securities class action questions whether the AI business can replace search.

Richard TangRichard Tangthree-minute read
Share
Baidu chairman and CEO Robin Li
1 / 5Slide 1 of 5
Robin Li, Baidu's chairman and chief executive. His company's new open-weights license turns free into paid at $50 million of deployer revenue.

Baidu+0.63% — Baidu, up 0.63 percent today published open weights for ERNIE-5.5-Flash-Base on November 5, and the interesting part is not the model 2. It is the license: free to use only below $50 million a year in related revenue, banned in the EU, UK and South Korea, and written by a company being sued for allegedly overstating how fast its AI business grows 2. Open weights with a revenue trigger are not open source; they are a billing instrument 2.

The model is real; the product is not

The checkpoint itself is real engineering: 280 billion total parameters with 14 billion active per token, pretrained on 33.5 trillion tokens, a 1 million-token inference window, and on Baidu's own harness it leads most rows against comparable base models 12.

But it is a base checkpoint, the end of pre-training. There is no instruct model, no hosted endpoint, no independent score, and per OrcaRouter's check on announcement morning, the primary repository had four downloads 2. A fine-tuning recipe ships too, validated only on Nvidia Blackwell GB200 and GB300 systems, chips export controls keep out of China 1. So the model that could start earning for Baidu does not exist yet, and nobody outside Baidu has confirmed the numbers that would make it worth adopting.

The trigger meets the lawsuit

Read the trigger against the litigation. A securities class action filed in September in the Southern District of New York alleges management overstated AI's ability to offset online marketing decay; in Q2 that ad revenue fell 19% year over year while AI Cloud Infra fell 17% quarter over quarter, and the ADS dropped 12.73% on the print 4. Lead-plaintiff motions are due November 13 4.

Every line of the Q2 print fell: ads, AI cloud and the stock

-0%-5%-10%-15%-20%Online marketing YoYAI Cloud Infra QoQADS one-day drop-12.73%-19%
Data
Decline
Online marketing YoY-19%
AI Cloud Infra QoQ-17%
ADS one-day drop-12.73%
Declines cited in the securities complaint: online marketing year over year and AI Cloud Infra quarter over quarter in Q2; the ADS move is the single-day drop on the earnings print. Source: PR Newswire via Morningstar.4

Baidu's CFO says AI margins are closing in on search's; the complaint's allegation says the cushion was overstated 34. A license that demands written authorization and fees once a deployer crosses $50 million is a bet that ERNIE gets used commercially at scale 34. It is also the only clause in this release that generates revenue Baidu can book.

The rails are built; the revenue is not

The same week shows the monetization stack assembling. TokenHub, Baidu's new marketplace, lists ERNIE beside DeepSeek, Kimi and GLM with unified billing across 398 million invocations in seven days 5.

The Kunlunxin chip unit's $2 billion Hong Kong IPO, meanwhile, slipped on pending regulatory clearance 6. Weights, marketplace, chips: the rails are built. The gap between them and revenue is the unfinished middle, and whoever deploys ERNIE past the $50 million line is the one doing Baidu's billing for it 6. Until a post-trained model ships and someone reproduces the benchmark table, shareholders are lending, not collecting.

Get Model by email, free

The AI labs, their models, and the business behind them. In our daily email, every weekday.

Loading BIDU

Reader comments

0 comments

    Model

    Anthropic hired a UK lobby firm the week it priced its $1.4 trillion IPO

    The same week Anthropic set its $70-to-78 IPO range, the UK lobbyist registry showed it paying for political access in three countries. Goldman's $685 billion revenue case runs on data centers that voters and regulators must permit, and buyers fund that insurance without seeing its cost.

    Model

    Mistral built its enterprise stack in eight days and priced none of it

    Between October 29 and November 5 the Paris lab published its biggest model's weights, expanded its agent platform and bought a fifth startup, all funded by a September mark of €21 billion. Not one piece of the stack carries a published price, while free Chinese open-weight models still outscore it.

    Model

    xAI's Grok 4.8 entered training 57 days late with no price card; Anthropic pays SpaceX $1.25 billion a month

    The 57-day wait bought a serving stack xAI wrote itself, and the delay ends with no scores and no price. For a SpaceX shareholder, the unpublished price card for 4.8 is the next receipt.

    Model

    Meta's Muse Ember costs $0.14 a task, a third of Claude Haiku, as Anthropic prices its IPO

    Meta's smallest model does the work agents repeat all day at $0.14 a task, matching Anthropic's Haiku at a third of the cost. Anthropic prices its IPO this week on about $81 billion of run-rate token revenue, and the floor under that business just moved.

    Model

    Anthropic cuts Fable 5.2 prices 20% on IPO roadshow eve, and Google and Broadcom shareholders carry the bill

    Anthropic discounted its flagship the day its IPO quiet period ended, for buyers who had already stopped climbing. The compute bill behind the discount is owed chiefly to Google and chipmaker Broadcom, which leaves their shareholders carrying the exposure.

    Model

    AfterQuery's agent training data lifted OpenAI's small model from 3.1% to 17% on Terminal-Bench

    The November 4 run, showcased on Thinking Machines' Tinker, sits beside rival Mercor's proof on the same page, giving labs a direct comparison of measured benchmark uplift for the first time. It is the claim beneath AfterQuery's reported $3.2 billion valuation, now competing in public for labs' data budgets.

    Get the newsletter

    The day's AI and market news, in your inbox every weekday.