Models · · 2 min read

Qwen 3.8 is a 2.4T-parameter model — but it's a preview API, not open weights yet

On July 19, 2026, Alibaba teased Qwen3.8-Max-Preview at 2.4 trillion parameters, claiming it's 'second only to Fable 5.' But the weights aren't out, there are no independent benchmarks, and today you can only reach it through the paid Token Plan. Here's what a builder should actually do with that.

Alibaba announced Qwen3.8-Max-Preview on July 19, 2026, a model it says has 2.4 trillion parameters. The Qwen team posted that the model “is launching and going open-weight soon” and called it “one of the most powerful models available today … second only to Fable 5.” (Source: Qwen on X)

Key facts:

  • Qwen3.8 has 2.4 trillion parameters. (Source: Qwen official announcement, 2026-07-19)
  • Today it exists only as Qwen3.8-Max-Preview, a proprietary preview. Open weights are promised “soon” with no date given. (Source: OfficeChai, 2026-07-19)
  • There are no independent benchmarks. The “second only to Fable 5” claim rests on Alibaba’s own internal evaluations — nothing from Artificial Analysis or LMArena yet. (Source: OfficeChai, 2026-07-19)
  • Access is through the Token Plan: a $6 Lite tier (2,500 credits / 7 days) up to a $68 Pro tier (40,000 credits / 7 days, 6–8 concurrent agents). (Source: OfficeChai, 2026-07-19)
  • The preview is served on Qwen Studio, Alibaba Cloud, and Qoder, and is compatible with both the OpenAI and Anthropic API protocols. (Source: OfficeChai, 2026-07-19)

The competitive framing is a trillion-parameter race between Chinese labs. Qwen3.8’s 2.4T sits just under Kimi K3’s 2.8T, and Alibaba owns roughly 36% of Moonshot AI, the lab behind Kimi — so this is partly a portfolio company competing with itself. (Source: OfficeChai, 2026-07-19)

Official Qwen 3.8 announcement graphic reading 'Qwen 3.8 is coming soon' with 2.4T and Open Weight badges
Alibaba's own announcement card is explicit that the model is "coming soon" — the 2.4T and Open Weight badges are a roadmap, not a download link. (Source: Qwen on X)

What this means if you’re a builder

The headline number is real, but three things are missing that a builder needs before betting on a model: open weights, independent benchmarks, and a stable model id you can self-host. Treat this as a signal, not a dependency.

  • Don’t rewrite anything for a 2.4T claim with zero third-party benchmarks. “Second only to Fable 5” is a vendor’s internal eval. Wait for Artificial Analysis / LMArena numbers, or run your own eval on the preview before it changes anything in your stack.
  • If you want open weights today, you already have options. For self-hosted coding you can run GLM-5.2 locally or use Qwen 3.6 for local coding; for a frontier open-weight model you can reach Kimi K3 now. Qwen3.8’s weights are a future event — the Qwen open-weight watch is the pattern here: Max-Preview ships first, open weights follow later.
  • The OpenAI/Anthropic API compatibility is the one thing you can use now. Because the preview speaks both protocols, testing it is a base_url + model-id swap — put the model id in config behind a flag so you can A/B it against your current model and drop in the open weights the day they land, without touching call sites. That’s the same portability discipline we cover in the Qwen 3.7 Max guide.

Bottom line: Qwen3.8-Max-Preview is worth a probe on the cheap Token Plan tier, but it is not something to build on until the weights and independent benchmarks show up.

Sources

Source: OfficeChai