Read time: ~5 minutes. Key facts:

• Alibaba previewed Qwen3.8-Max on July 19, 2026 — a claimed 2.4 trillion parameters, described as the first multimodal model above 1T params (text, images, video, documents).
• It’s a preview. The active-parameter count, benchmarks, license, and open-weights date are all undisclosed — and there’s no HuggingFace repo yet, so you can’t run it locally today.
• Access now: via Token Plan, Coder, and CoderWork, at 10% of the standard rate, over OpenAI- and Anthropic-compatible API protocols.
• Alibaba says open weights are coming “soon” with no date — which would break its pattern of shipping the last two Max flagships closed.

Sourcing note: all figures are Alibaba’s own preview claims, reported by outlets covering the July 19 announcement (MarkTechPost and others). Where a number isn’t public, this guide says so rather than guessing. A “run Qwen 3.8 locally” guide will follow once the weights and a model card actually ship. Links at the bottom.

Days after Moonshot open-sourced Kimi K3, Alibaba fired back with a preview of Qwen3.8-Max — a 2.4-trillion-parameter multimodal model it positions just behind Claude Fable 5. It’s a big claim. It’s also, right now, a preview with more blanks than a spec sheet should have — so the honest question isn’t “how do I run it,” it’s “what do we actually know, and how do I try it today.” Here’s both.

---

1. What’s actually confirmed

From Alibaba’s preview announcement:

• Qwen3.8-Max-Preview is the official name.
• 2.4 trillion total parameters (Alibaba’s claim), making it — per Alibaba — the first multimodal model above 1 trillion parameters.
• Multimodal: it handles text, images, video, and documents.
• Positioned “second only to Fable 5” among the systems Alibaba evaluated — i.e. a claimed near-frontier model.
• It’s available now as a preview (see §3), and Alibaba says an open-weight release is coming “soon.”

That last point is genuinely notable: Alibaba’s last two Max flagships shipped closed. An open-weight Max would break that pattern and put a 2.4T multimodal model into the same self-hostable category as Kimi K3 and Inkling.

---

2. What Alibaba hasn’t disclosed (the part that matters)

Here’s where to keep your expectations calibrated. As of the preview, these are not public:

• Active parameter count. For a sparse Mixture-of-Experts model, this is the number that determines real serving cost — how much actually fires per token. “2.4T total” tells you almost nothing about what it costs to run without it. Alibaba hasn’t said.
• Benchmarks. No benchmark table has been published. The “second only to Fable 5” ranking rests on internal evaluation — treat it as a marketing claim until third-party numbers exist.
• License. Not published. “Open weights soon” with no license text means you can’t yet know the terms you’d be agreeing to.
• Open-weights date. Promised “soon,” no specific date.
• A HuggingFace repo / model card. None yet — which is why local inference (vLLM, GGUF, llama.cpp) isn’t possible today.

None of this makes Qwen 3.8 vaporware — it’s a real, accessible preview. But a responsible read is: strong claims, thin disclosure. Wait for the active-param count and independent benchmarks before betting a roadmap on it.

---

3. How to access the Max Preview today

You can use Qwen3.8-Max-Preview right now — just not locally. From the announcement:

• It’s available through Token Plan, Coder, and CoderWork (Alibaba’s Qwen access tiers).
• Pricing: 10% of the standard rate during the preview — an aggressive discount to drive adoption while it’s in preview.
• Protocol compatibility: OpenAI- and Anthropic-compatible APIs — so, like other Qwen Max models, you can point existing OpenAI- or Anthropic-SDK tooling at it by swapping the base URL and model name, rather than rewriting your client.

Practically: if you already build against OpenAI or Anthropic SDKs, trying Qwen3.8-Max-Preview is a base-URL-plus-model-name change through one of the Qwen plans — and at 10% of standard pricing, it’s cheap to evaluate. Just remember you’re testing a preview whose behavior (and price) can change before GA.

---

4. Running it locally: not yet — here’s the trigger

To be clear: there is no way to run Qwen 3.8 locally as of this writing. No open weights, no HuggingFace repo, no GGUF, no vLLM recipe. The blockers are simply that the weights haven’t shipped.

When they do — Alibaba’s “soon,” no date attached — expect the same honest-hardware math as every other giant model: a 2.4T model, even quantized, needs serious memory to hold, so local will be high-RAM-workstation / multi-GPU territory, not a laptop. (For the shape of that, see how it played out for Inkling (975B) and the access-first approach we took for Kimi K3, which is in the same “weights coming later” boat.)

We’ll publish a “run Qwen 3.8 locally” guide the moment the weights, license, and a real vLLM/GGUF path are out and verifiable — no point writing local commands against a model with no download.

The takeaway

Qwen3.8-Max-Preview is a real, accessible preview of a claimed 2.4T-parameter, first-of-its-kind >1T multimodal model — and it’s cheap to try today (Token Plan / Coder, 10% of standard rate, OpenAI/Anthropic-compatible). But it’s a preview in the fullest sense: active params, benchmarks, license, and the open-weights date are all still blank, and there’s no local path yet. Use the preview to form your own opinion, take the “second only to Fable 5” claim as unverified until benchmarks land, and watch for the weights drop — at which point the run-locally story (and our guide) begins.

For models you can actually self-host or that are further along, see Kimi K3 (access + weights July 27), Qwen 3.6 for local coding, and running Inkling locally.

Sources

• Alibaba Previews Qwen3.8-Max, a 2.4T-Parameter Multimodal Model — MarkTechPost — Qwen3.8-Max-Preview, 2.4T total params (active count undisclosed), first multimodal model above 1T params (text/images/video/documents), no benchmark table published (“second only to Fable 5” is internal/unverified), access via Token Plan / Coder / CoderWork at 10% of standard rate, OpenAI- and Anthropic-compatible protocols, open weights “soon” with no date / no license / no HuggingFace repo, breaking the closed-flagship pattern of the last two Max models
• Alibaba Launches Qwen 3.8 With 2.4 Trillion Parameters — MLQ — near-frontier performance claim; active-parameter count highlighted as the key undisclosed detail
• All performance figures are Alibaba’s own preview claims, pending independent benchmarks and a weights/license release. Verified July 20, 2026 — a local-run guide will follow once the model card ships.