China AI Bench
5 min read

Qwen 3.8: Alibaba's 2.4T-parameter launch, one week in

ReviewBenchmarks Deskby Benchmarks Desk, Testing & rankings desk

One week after Alibaba's Qwen 3.8 launch: what we can verify (2.4T params, ~95B active, IDC Leaders quadrant, eight day-one integrations) and what we can't (API pricing, independent benchmarks, the weights).

A 2.4-trillion-parameter model shipped last week, and the most striking thing about the launch is what Alibaba didn't publish alongside it: a benchmark table, an API price, or the weights.

That's unusual for a release this size, and it makes Qwen 3.8 an interesting test for how we run this site. We rank what we can source. This release has specs we can source, claims we can label, and a pricing column we have to mark n/r.

The launch

Alibaba released Qwen 3.8 on August 3, 2026, with the Max-Preview build live immediately on its Token Plan, Qoder, and QoderWork products [1]. It's the first Qwen model above a trillion parameters: 2.4 trillion total, roughly 95B active per token, MoE, and multimodal [1]. Alibaba positions it for long-horizon work: coding, office agent tasks, data analysis, and Office workflows, with a claimed step up over Qwen 3.7-Max in code engineering and office capability [1][2].

The company says open weights are coming "next week" [1][2]. As of August 6 they aren't out, so that line stays in the vendor column until it lands.

What we can verify

Three things are independently checkable today.

IDC puts Alibaba Cloud in the Leaders quadrant. The IDC MarketScape Worldwide Foundation Model Software 2026 report lists 11 vendors; Alibaba Cloud is the only Chinese vendor in the Leaders quadrant, alongside Google, Anthropic, OpenAI, and Amazon [5]. An analyst firm's map of the market, not a score, but a third-party signal that the Qwen family reads as enterprise-grade outside China.

Day-one integrations happened, and they were fast. On the day coverage went out, OpenRouter, OpenCode, Hermes Agent, Command Code, Vercel, Novita, Charm, and DeepInfra all announced Qwen 3.8 support [4]. That covers API aggregators, coding agents, and hosting platforms. For a model whose weights aren't public yet, this is the ecosystem betting on the brand and the API.

The consumer price is out, the API price is not. Qoder's promo runs 14 days of Pro trial with 300 credits, daytime credits as low as 10% off, and a ¥35/month entry tier [3]. API pricing per million tokens is n/r: Alibaba hadn't published it by August 6, and we don't fill that cell with estimates.

Also worth noting: the same week brought Qwen 3.0 Image Pro, an image generation model that hit the Hacker News front page [6]. Alibaba is shipping on two tracks at once, base model and image model, and both are getting Western developer attention.

The vendor claim we're flagging

Alibaba says Qwen 3.8's overall capability is "second only to Claude Fable 5" [1][3]. That line is vendor-reported. There is no independent benchmark behind it as of this writing: no Artificial Analysis intelligence score, no LMArena placement, no CAISI eval. We label it as what it is and move on. When an independent number appears, we will update this piece and the leaderboard row.

For context on the same board: Kimi K3 sits at 57.1 on the AA Intelligence Index and took the Frontend Code Arena top spot with 1679 [7][8]; GLM-5.2 holds the open-source lead on AA at 51 [9]. Qwen 3.8 currently has no comparable number. That's the honest gap in this launch.

Our take

Alibaba is running a spec-first launch: announce the biggest Qwen ever, get the ecosystem to plug in, let the market price it. It works because day-one support is now table stakes for open-weight families. OpenRouter, OpenCode, and Vercel added K3 within hours of its release in July [10]; they did the same for Qwen 3.8 without weights and without a benchmark. The distribution layer doesn't wait for proof anymore.

Here's what gets me: a release this big with this little data is itself a positioning statement. The claim "second only to Fable 5" with no link is marketing, and Alibaba knows we know. The IDC quadrant is the more interesting card, because it says "enterprise procurement" rather than "benchmark heroics." Qwen's real competition isn't the frontier labs, it's Kimi and GLM for the open-weight middle class of developers.

Risks and what we're watching

Three things can change this story fast.

  • The weights. "Next week" was said on August 3. If the release slips, the day-one goodwill erodes. We're tracking the HF repo.
  • API pricing. Qwen 3.7-Max pricing put output at ¥36 per million tokens as of late July [11]. Where 3.8 lands decides the fight: K3 territory ($3/$15 [12]) or GLM territory ($1.4/$4.4 [13]). Those are different products.
  • Independent evals. The first AA or Arena number will either back the "second only to Fable 5" line or quietly bury it.

Bottom line

Qwen 3.8 is, so far, a spec sheet and an ecosystem signal. Alibaba is a Leader in an analyst quadrant, its API got plugged into eight platforms in a day, and its flagship model has no public benchmark and no API price. We rank what we can source, so the leaderboard row for Qwen 3.8 says exactly that.

Same number on our leaderboard → Qwen 3.8 row

Sources

  1. Tencent News (Alibaba Qwen 3.8 release report), news.qq.com/rain/a/20260804A0ACRT00 (published 2026-08-03, captured 2026-08-06, media report of vendor announcement)
  2. ifeng Tech (Qwen 3.8-Max details), tech.ifeng.com/c/8vHhZ7YlyBr (published 2026-08-03, captured 2026-08-06, media report)
  3. Sina Finance / Zhitong (pre-announcement, vendor self-assessment, Qoder promo), finance.sina.com.cn/stock/hkstock/ggscyd/2026-07-20/doc-iniimuyk5025993.shtml (published 2026-07-20, captured 2026-08-06, media report)
  4. Leiphone (day-one integrations), leiphone.com/category/industrynews/Ufuyq9sSvBlkp2Ix.html (published 2026-08-04, captured 2026-08-06, media report)
  5. Leiphone (IDC MarketScape 2026), leiphone.com/category/industrynews/1gS0E0CcUDxEgbQR.html (published 2026-08-05, captured 2026-08-06, third-party eval, via media)
  6. Qwen Cloud (Qwen 3.0 Image Pro model page), qwencloud.com/models/qwen-image-3.0-pro (published 2026-08-05, captured 2026-08-06, official)
  7. MACGPU Blog (K3 AA Intelligence Index 57.1, #3/189), macgpu.com/zh/blog/2026-0722-kimi-k3-kaiyuan-quanzhong-fabu.html (published 2026-07-22, captured 2026-08-06, third-party eval, via community blog)
  8. Jiemian (K3 Frontend Code Arena 1679), jiemian.com/article/14786633.html (published 2026-07-17, captured 2026-08-06, third-party eval, via media)
  9. Juejin (GLM-5.2 AA 51 open-source SOTA), juejin.cn/post/7654122741078671402 (published 2026-06-24, captured 2026-08-06, third-party eval, via community)
  10. Leiphone (K3 day-one ecosystem), huacheng.gz-cmc.com/pages/2026/07/28/6370587b2a104859af17da3aa866fc36.html (published 2026-07-28, captured 2026-08-06, media report)
  11. Qwen 3.7-Max output pricing ¥36/MTok, Morgan Stanley research compiled by 观点网 (snapshot 2026-07-22, captured 2026-08-06 [media report]
  12. Ollama (K3 API pricing $3/$15, cache hit $0.30), ollama.com/library/kimi-k3 (published 2026-08-01, captured 2026-08-06, official)
  13. 36Kr (GLM-5.2 API pricing $1.4/$4.4), 36kr.com/p/3919319636290946 (published 2026-07-31, captured 2026-08-06, official pricing, via media)

Disclosure: no sponsors, no affiliate links. All numbers above are compiled from public sources with capture dates; we did not run Qwen 3.8 ourselves. Vendor-reported claims are labeled as such.