Alibaba Self-Ranks AI, Skips Benchmarks.

Mazen Al Sabbagh

— 5 min read
Share

TL;DR

  • Qwen3.8-Max: Alibaba unveiled a 2.4-trillion-parameter multimodal model at WAIC, asserting its capability as "second only to Fable 5" based solely on internal metrics. No public benchmarks, model card, or weights accompanied this claim.
  • Moonshot: New Kimi subscriptions were halted due to K3 demand nearing GPU capacity. Bloomberg reports a Hong Kong IPO is targeted within six months, potentially valuing the company above $30 billion.
  • Fable 5: The deadline for Fable 5's subscription status was resolved preemptively: permanent access, at half limits, begins Monday for Max and Team Premium users. Pro users receive a one-time $100 credit.
  • Data centers: A national anti-data-center mobilization, organized by HumansFirst, saw 142 protests across 42 states on Saturday, marking a significant escalation in infrastructure opposition.
  • Washington: Discussions continued over the weekend regarding a proposed FINRA-style body, developed with Treasury Secretary Bessent's input, to vet frontier models pre-release, keeping the regulatory debate prominent.

Lead Story: Alibaba Self-Ranks AI, Skips Benchmarks

Alibaba leveraged day three of the World Artificial Intelligence Conference in Shanghai to debut Qwen3.8-Max-Preview, a 2.4-trillion-parameter sparse mixture-of-experts model capable across text, images, video, and documents. This marks Alibaba's first Qwen multimodal entry above the 1-trillion-parameter threshold, and the company's assessment is unequivocal: Qwen3.8-Max is positioned as "one of the most powerful models available today... second only to Fable 5," surpassing prominent models from OpenAI and Google.

This assertive ranking, however, is critically qualified. It relies entirely on Alibaba's internal evaluations, conspicuously absent of any public benchmark tables, model cards, active-parameter counts, or licensing details. While open weights are vaguely promised "soon," current access remains restricted to Alibaba's Token Plan subscription and its Qoder coding tools, offered at 10% of standard pricing during this preview phase.

The timing of this announcement is the compelling narrative. Qwen3.8-Max represents the third Chinese frontier claim this week alone, following Moonshot's Kimi K3 debut on Thursday as the largest open model, and preceding the widely anticipated DeepSeek V4. After K3's independent performance placed it fourth globally and Friday's market selloff confirmed investor recognition of Chinese labs, Alibaba's unvalidated claim will face immediate scrutiny. Until verifiable benchmarks emerge, "second only to Fable 5" functions as a press strategy, not a validated outcome.

In Other News

Moonshot reaches capacity limits — while pursuing a $30 billion IPO. Moonshot today halted new Kimi subscriptions, citing "unprecedented compute challenges" from post-K3 launch demand, pushing its clusters to their operational boundaries. Concurrently, Bloomberg reports the company informed investors of a planned Hong Kong listing within six months, eyeing a valuation above $30 billion from a closing funding round. Underwriter discussions with Goldman Sachs and CICC are underway, alongside efforts to dismantle its offshore VIE structure. Annual recurring revenue hit $300 million in June, up from $200 million in April. A lab rejecting paying customers merely three days post-launch offers the most definitive demand signal yet from China's AI sector.

Fable 5's subscription deadline resolves proactively. Today marked the thrice-extended deadline for Fable 5's tenure within Claude subscription plans. Anthropic preempted this by confirming permanent inclusion in Max and Team Premium tiers, subject to 50% of weekly usage limits, effective Monday. Pro and Team Standard users will lose bundled access, receiving a one-time $100 credit, with Fable thereafter relying on prepaid API credits at $10/$50 per million tokens. The temporary 50% Claude Code limit boost also concludes Monday, ensuring substantial compute constraint for power users even on Max.

The data-center opposition expands nationally. Opponents of AI infrastructure development organized 142 protests across 42 states on Saturday, representing the first coordinated national mobilization spearheaded by HumansFirst. Texas led with 18 demonstrations, while Georgia hosted 11. The group's demands are notably targeted—transparent approvals, environmental protections, community benefits, and binding developer accountability—rather than advocating outright abolition. While turnout in some cities was modest, the established organizational framework introduces a new variable into every hyperscaler's site selection calculus.

A FINRA-like body for frontier AI garners weekend attention. The Trump administration's reported proposal for an independent entity under SEC supervision to vet frontier models—involving voluntary 30-day pre-release reviews for cyber, bio, and deception risk—continued its policy discourse, with a Forbes op-ed today suggesting an industry-funded body could attract talent inaccessible to traditional agencies. This constitutes a government-supervised analogue to the industry-led standards body Demis Hassabis proposed last week, with Treasury Secretary Scott Bessent reportedly contributing to its development.

X / Social Pulse

The K3 discourse dominated the weekend: Arena chief Anastasios Angelopoulos called it potentially "the single biggest release of the year" and the moment Chinese open models eclipsed US counterparts, while Ethan Mollick deemed it "closest to the frontier yet," albeit noting its deficiency in writing a "good murder mystery." Simon Willison declared "no Fablepocalypse," contending GPT-5.6 Sol's competition made Fable 5's complete removal untenable for Anthropic. Moonshot's subscription-pause communique ignited Hacker News, interpreted as a scarcity play veiled as an apology. And Elon Musk alluded that the 2-trillion-parameter Grok 4.6 finishes initial training "next week" and "might be able to exceed Kimi" — the frontier race now openly benchmarks against a Chinese open model.

One to Watch

Monday's Seoul open. Korean markets remained closed Friday for Constitution Day, precluding Samsung and SK Hynix from trading on the K3-driven global selloff that saw the Nasdaq decline 2.9% for the week. Monday also brings WAIC's conclusion, the Fable 5 tier restructuring, and continued anticipation for DeepSeek V4.

Quick Hits

  • Market Performance: Friday concluded a challenging trading week on Wall Street, with the Nasdaq down 2.9% and the S&P 500 down 1.6% for the week. The SMH chip ETF declined almost 9%, with Apple reclaiming the top market-cap ranking from Nvidia.
  • Humanoid Production: China's humanoid-robot output surpassed 40,000 units in H1 and is projected to exceed 100,000 this year, with AgiBot's co-founder advocating for China to "export its technologies globally."
  • Cyber Capabilities: The UK's AI Security Institute reported open-weight models now achieve offensive-cyber capabilities matching closed-frontier models from only 4-7 months prior, a acceleration from the 6-10 month lag observed through 2025.
  • Microsoft's Initiative: Microsoft is reportedly preparing "Project Perception," an AI security scanner akin to Mythos, slated for launch as early as this month. It routes across Anthropic, OpenAI, and internal models to competitively price against Anthropic.
  • Pricing Strategy: Alibaba's preview pricing gambit—10% of standard rates—extends the Chinese price war that K3's $15-per-million output tokens had appeared to conclude.

A week that commenced with a singular Chinese frontier claim now concludes with three, while the only discernible US market response was a subscription-tier retrenchment and a selloff. Monday's Seoul open and the DeepSeek V4 watch will indicate whether markets perceive this gap as substantive or the panic as overextended.

Sources

Lock in. M. mazen@thorterminal.com

Read more