Alibaba Unveils Qwen3.8 at WAIC: A 2.4-Trillion-Parameter Frontier Model — and a Promise to Open-Source It

`json { "skip_run": false, "title": "Alibaba Unveils Qwen3.8 at WAIC: A 2.4-Trillion-Parameter Frontier Model — and a Promise to Open-Source It", "excerpt": "Announced at the World Artificial Intelligence Conference in Shanghai, Alibaba's Qwen3.8 is a 2.4-trillion-parameter multimodal MoE model that the company claims is 'second only to Fable 5' — and, in a break from precedent, Alibaba has pledged to release the weights as open-source. The preview is live now through Token Plan, Qoder, and QoderWork at 10% of standard pricing.", "tags": ["Qwen", "Alibaba", "Qwen3.8", "Open-Weight", "China AI", "WAIC 2026", "MoE", "Multimodal", "Frontier Models", "Developer Tools", "Apple Intelligence"], "sources": [ {"title": "Alibaba previews Qwen3.8, claims it's second only to Claude Fable 5 — SiliconAngle", "url": "siliconangle.com↗"}, {"title": "Alibaba Launches Qwen 3.8 With 2.4 Trillion Parameters — MLQ.ai", "url": "mlq.ai↗"}, {"title": "Qwen3.8-Max Preview: Token Plan Pricing and Open Weights Soon — ExplainX.ai", "url": "explainx.ai↗"}, {"title": "Qwen3.8 Preview: 2.4T Params, Open Weights Release — BuildFastWithAI", "url": "buildfastwithai.com↗"}, {"title": "Alibaba's Qwen takes on Kimi K3 with open-weight Qwen 3.8 — The Decoder", "url": "the-decoder.com↗"}, {"title": "Qwen3.8 Official Announcement — Alibaba Qwen on X", "url": "x.com↗"}, {"title": "Apple Intelligence approved for launch in China with Alibaba's Qwen AI — TechCrunch", "url": "techcrunch.com↗"}, {"title": "Apple Intelligence wins China approval after 22 months — TechTimes", "url": "techtimes.com↗"}, {"title": "Alibaba Announces 2.4 Trillion-Parameter Open-Weight Qwen 3.8 — OfficeChai", "url": "officechai.com↗"}, {"title": "WAIC 2026 Shanghai — AI Conference Overview", "url": "aiii.global↗"}, {"title": "Qwen3.8 GoldieBench evaluation — GoldieBench", "url": "goldiebench.com↗"}, {"title": "DeepSeek V4 API Pricing — DeepSeek API Docs", "url": "api-docs.deepseek.com↗"}, {"title": "Alibaba Cloud Model Studio Pricing — Alibaba Cloud", "url": "help.aliyun.com↗"}, {"title": "Qwen API Platform — Alibaba Qwen", "url": "qwen.ai↗"}, {"title": "Qwen3.8 on Hugging Face — Qwen org", "url": "huggingface.co↗"}, {"title": "DeepSeek-V4-Pro Model Weights — Hugging Face", "url": "huggingface.co↗"} ] } ```

Alibaba Unveils Qwen3.8 at WAIC: A 2.4-Trillion-Parameter Frontier Model — and a Promise to Open-Source It

On the final day of the World Artificial Intelligence Conference in Shanghai — a four-day event that drew over 1,400 international guests, 1,100 enterprises, and a rare appearance by President Xi Jinping — Alibaba chose the moment to drop its biggest model yet. Qwen3.8, announced on July 19, 2026, is a 2.4-trillion-parameter multimodal Mixture-of-Experts model that the company claims is "second only to Fable 5," Anthropic's current flagship. More significantly, Alibaba has pledged to release the weights as open-source — a departure from its established practice of keeping Max-tier models closed.

The preview, called Qwen3.8-Max-Preview, is already accessible through Alibaba's Token Plan subscription, its Qoder coding platform, and QoderWork productivity suite, priced at 10% of standard rates during the launch period. Full open weights are promised "soon," with no specific date or license terms yet published.

"Qwen3.8 is launching and going open-weight soon. 2.4T parameters, continuously evolving. One of the most powerful models available today, compatible to leading frontier AI models, second only to Fable 5." > — Alibaba Qwen team, official X post↗, July 19, 2026

The timing is deliberate. Moonshot AI's Kimi K3 — a 2.8-trillion-parameter open-weight model — launched just three days earlier and immediately overwhelmed Moonshot's own infrastructure, forcing the company to suspend new consumer subscriptions. Alibaba is positioning Qwen3.8 as the stable, enterprise-ready alternative in the same parameter tier, backed by the cloud infrastructure of one of China's largest technology companies.

What Qwen3.8 Actually Is

Qwen3.8 is the first model in the Qwen family to exceed one trillion parameters while also being multimodal. It processes text, images, video, and documents within a single architecture — a capability its predecessor, Qwen3.7-Max, did not offer at that scale. The model uses a sparse MoE design, meaning only a fraction of its 2.4 trillion parameters are activated per inference call, though Alibaba has not disclosed the activated-parameter count — a critical omission for anyone planning self-hosted deployment.

The SiliconAngle report↗ from the WAIC announcement notes that Alibaba's previous flagship, Qwen3.7-Max, shipped in May with a full set of published results, including a score of 56.6 on the Artificial Analysis Intelligence Index. Qwen3.8 arrived with no model card, no benchmark table, and no activated-parameter specification. The company described it only as "continuously evolving" — language that suggests rolling updates to the preview checkpoint rather than a fixed, versioned release.

Key Specifications (as disclosed)

• Total parameters: 2.4 trillion (MoE architecture)
• Activated parameters per token: Not disclosed
• Modalities: Text, images, video, documents
• Context window: Not yet specified for the 2.4T variant
• Current access: Qwen3.8-Max-Preview via Token Plan, Qoder, QoderWork
• Open-weight status: Promised "soon" — no date, no license terms published
• Preview pricing: 10% of standard rates during launch period

For context, Qwen3.7-Max — the model this replaces at the top of the Qwen lineup — scored 92.4% on GPQA Diamond, 80.4% on SWE-bench Verified, and 69.7% on Terminal-Bench 2.0 in Alibaba's May benchmarks. Those numbers established Qwen3.7-Max as a credible frontier competitor. Whether Qwen3.8 improves on them remains unverified.

The Open-Weight Pledge: Why It Matters

The most consequential part of the Qwen3.8 announcement is not the parameter count — it is the open-weight commitment. Alibaba's Max-tier models have historically remained closed-source. Qwen3.7-Max, Qwen3-Max, and their predecessors were available only through Alibaba's API. The promise to release Qwen3.8 weights would mark the first time a Qwen flagship has been made downloadable.

If Alibaba follows through, Qwen3.8 would become the largest open-weight model ever released — surpassing even Kimi K3's 2.8-trillion-parameter offering in strategic significance, given Alibaba's broader developer ecosystem and the Apache 2.0 licensing that has characterized the rest of the Qwen family.

The practical implications are significant. An open-weight Qwen3.8 would allow:

• Self-hosted deployment for enterprises with data-sovereignty requirements
• Fine-tuning on proprietary datasets without routing data through Alibaba's cloud
• Air-gapped operation for regulated industries and government customers
• Community-driven quantization — the same GGUF and GPTQ variants that made Qwen3.6-27B a popular local model

However, the caveat is real: at 2.4 trillion total parameters, even a heavily quantized Qwen3.8 will require multi-node inference infrastructure. The Qwen3.6-27B remains the practical local option for teams without data-center-scale hardware. Expect distilled smaller variants — likely in the 30B-70B range — to follow the flagship release, as has been Alibaba's pattern with previous Qwen generations.

How to Access It Today

The preview is live. Developers can reach Qwen3.8-Max-Preview through Alibaba's Token Plan↗, which bundles multiple frontier models under a single subscription. The plan includes not just Qwen3.8-Max-Preview but also GLM-5.2, DeepSeek-V4-Pro, and Wan2.7-image-pro — making it one of the more unusual subscription offerings in the current market, where a single monthly fee buys access to models from competing Chinese labs.

Token Plan Pricing (Individual, July 2026 launch discounts)

• Lite: $6/month (discounted from $8) — 2,500 weekly credits, 700 per 5-hour burst window, 1–2 concurrent agents
• Standard: $18/month (discounted from $25) — 10,000 weekly credits, 3,000 per burst, 3–4 agents
• Pro: $68/month (discounted from $80) — 40,000 weekly credits, 12,000 per burst, 6–8 agents

The plan exposes both OpenAI-compatible and Anthropic-compatible API endpoints, meaning developers can route Qwen3.8-Max-Preview through Claude Code, Cursor, OpenCode, Cline, or any harness that accepts a custom base URL. International access is through `qwencloud.com`; mainland China users access `platform.qianwenai.com`. The two regions use separate accounts and billing.

For standalone API access without a subscription, Alibaba has not yet published per-token pricing for Qwen3.8. The predecessor Qwen3.7-Max was priced at $1.25 per million input tokens and $3.75 per million output tokens — a useful baseline for budget planning until official Qwen3.8 rates appear. The Alibaba Cloud pricing documentation↗ covers the existing Qwen catalog; Qwen3.8 pricing will appear there once the model exits preview.

The Apple Intelligence Connection

The Qwen3.8 announcement lands in the same week as another major Alibaba AI milestone. On July 15, 2026, the Cyberspace Administration of China approved Apple Intelligence for deployment in mainland China — ending a 22-month regulatory delay — with Alibaba's Qwen serving as the primary language AI engine for the Chinese version of the service.

According to TechCrunch's reporting↗, the Chinese Apple Intelligence implementation uses a dual-partner architecture: Qwen handles text and image understanding and generation, while Baidu handles visual search and AI-powered search. The arrangement gives Alibaba distribution reach into Apple's massive Chinese user base — an estimated 50+ million active iPhone users — without requiring Apple to build its own Chinese-language foundation model.

The engineering challenge is substantial. Reports indicate that Alibaba's Qwen3.6-27B — which typically occupies 54 GB at standard precision — has been compressed to under 4 GB for on-device operation on iPhone hardware. Whether this compression uses 1-bit quantization or another technique has not been confirmed. The privacy architecture also differs from Apple's global Private Cloud Compute system, which was designed for Apple's own foundation models and is unlikely to extend to third-party Chinese providers.

For Alibaba, the Apple deal is a commercial validation that complements the Qwen3.8 announcement. It demonstrates that Qwen models are production-grade at consumer scale, not just benchmark performers.

Competitive Landscape: Where Qwen3.8 Sits

The Chinese frontier model market has compressed dramatically in the past month. Three models now occupy the multi-trillion-parameter tier:

• Kimi K3 (Moonshot AI): 2.8 trillion parameters, open-weight under Modified MIT license, weights dropping July 27. Debuted at #1 on the Frontend Code Arena. Already overwhelmed Moonshot's infrastructure.
• Qwen3.8 (Alibaba): 2.4 trillion parameters, multimodal, open-weight promised. Preview available now. No published benchmarks yet.
• DeepSeek-V4-Pro (DeepSeek): 1.6 trillion parameters, open-weight under DeepSeek's custom license, available on Hugging Face↗. API pricing at $0.435/M input tokens (cache miss) and $0.87/M output tokens. Legacy model identifiers (`deepseek-chat`, `deepseek-reasoner`) retire on July 24, 2026 at 15:59 UTC.

Against Western models, Alibaba's "second only to Fable 5" claim positions Qwen3.8 above GPT-5.6 Sol, Gemini 3.1 Ultra, and every other frontier model except Anthropic's current leader. That is an extraordinary claim, and the absence of any benchmark data to support it is the most significant gap in the announcement.

The GoldieBench evaluation↗ — one of the few independent assessments available — places Qwen3.8 fourth among 18 frontier models with an average score of 8.16/10, showing particular strength in "Visuals" (8.7/10) and "Sims" (8.6/10) categories. That ranking puts it behind GPT-5.6 Sol and Fusion but ahead of Kimi K3 and Claude Fable 5 in those specific task categories. It is a single data point, not a comprehensive evaluation, but it suggests the frontier-performance claim is not entirely without basis.

What Developers Should Do Now

• Start a Token Plan Lite trial ($6/month) and run Qwen3.8-Max-Preview against your top coding and reasoning tasks before committing to a higher tier.
• Replace DeepSeek legacy model identifiers (`deepseek-chat`, `deepseek-reasoner`) before the July 24 deadline — this is urgent and unrelated to Qwen3.8.
• Do not redesign production infrastructure around Qwen3.8 until Alibaba publishes the model card, open weights, license terms, and standard per-token pricing.
• Watch the Hugging Face Qwen organization for the open-weight release; community quantizers (Unsloth, bartowski) typically publish GGUF variants within days of an official weight drop.
• Benchmark against Qwen3.7-Max as the control — it has published scores and stable API pricing, making it the right baseline for measuring whether Qwen3.8 actually improves on your workloads.

The WAIC Context: China's AI Moment

The choice to announce Qwen3.8 at WAIC 2026 was not accidental. The conference — held July 17–20 in Shanghai under the theme "Intelligent Partners, Co-create the Future" — was the largest in the event's history, with over 300 global product debuts and a formal address by President Xi Jinping calling for international AI cooperation and an end to "overstretching the national security concept" to justify chip export controls.

In that context, Qwen3.8's open-weight pledge carries a geopolitical dimension beyond its technical significance. China's AI labs have made open-weight releases a strategic differentiator against U.S. labs that have moved toward closed, API-only models. Alibaba releasing a 2.4-trillion-parameter model under an open license — if it follows through — would be the most visible demonstration yet of that strategy at frontier scale.

The week's events, taken together, tell a coherent story: Alibaba is positioning itself as China's default AI infrastructure provider, with Qwen models powering Apple's Chinese users, anchoring the Token Plan subscription ecosystem, and — if the open-weight promise holds — becoming the largest freely downloadable model in the world. Whether Qwen3.8 delivers on the "second only to Fable 5" claim is a question that benchmarks will answer. Whether Alibaba delivers on the open-weight promise is a question that the next few weeks will settle.

The preview is available now. The weights are not. Plan accordingly.