15% off all models 🎉 Every model at 85% of OpenRouter list price.Browse models →
Blog
ApiFlux Team6 min read

Qwen3.8-Max: 2.4T Parameters, Zero Benchmarks — What's Confirmed, What Isn't, and How to Test It Yourself

Alibaba's Qwen3.8-Max is live in preview: 2.4T parameters, multimodal, and — as of writing — no published benchmarks, no model card, and an open-weight date that has already slipped. A neutral gateway's read of confirmed vs. claimed, plus how to A/B qwen3.8-max against Claude, GPT, Gemini and DeepSeek on one key.

Qwen3.8-Max: 2.4T Parameters, Zero Benchmarks — What's Confirmed, What Isn't, and How to Test It Yourself

Short version (as of August 1, 2026): Alibaba announced Qwen3.8-Max-Preview on July 19 at WAIC in Shanghai. It's a 2.4-trillion-parameter multimodal model, callable today as a hosted preview through Alibaba's Token Plan. What it does not have yet: a single published benchmark, a model card, a disclosed context window, or the open weights the team promised. We route Qwen next to a dozen other models and have no stake in which one wins — so here's the plain read of what's confirmed vs. what's still a launch-stage claim, and how to decide for yourself instead of trusting a slide.

Qwen3.8-Max confirmed vs. claimed vs. missing — a three-column breakdown of the public record against the launch materials What's actually on the record vs. what's still just a claim. Sources at the bottom of this post.

What's actually confirmed

These are the facts multiple independent outlets reported from the July 19 announcement:

TopicConfirmed
ModelQwen3.8-Max-Preview — a preview, not a final release
Parameters2.4T total (sparse MoE; activated count not disclosed)
ModalityMultimodal: text, image, video, document — the first Qwen above 1T params to do this
AvailabilityHosted preview via API, on Alibaba's Token Plan (and Qoder / QoderWork)
AnnouncedJuly 19, 2026, at WAIC Shanghai
Self-positioningAlibaba's own evaluation puts it "second only to Claude Fable 5"

That's the honest extent of the confirmed record. Everything past this line is either a claim from the launch materials or a gap.

What's claimed but not verifiable

The launch brief and various write-ups carry a lot of impressive specifics. As a gateway, our job is to tell you which ones you can lean on — and right now, these you can't:

  • "1M-token context." It shows up in the launch brief, but Alibaba did not publish an official context-window spec. Treat the number as marketing until it's in the docs.
  • The "16-day autonomous build" (oh-my-cli). The brief describes Qwen3.8-Max autonomously building an open-source agent system with a public GitHub trace (265 commits, 127 PRs, 151 issues) at qwen-code-dev-bot/oh-my-cli. We could not locate that repository. A "public, verifiable trace" you can't open isn't verifiable — so we're not counting it. If Alibaba publishes the real link, we'll update this post.
  • The case studies — chip design over 500 iterations, a 365-day e-commerce simulation turning ¥100k into ¥416k, a first-place NVIDIA kernel benchmark, a paper reproduction beating its source on AIME24. These are vendor demos with no reproducible artifacts released. Interesting as hypotheses, not as evidence.
  • "Top tier globally." Standard launch-day self-assessment. See the next section for why it's currently unfalsifiable.

None of this means the claims are false. It means they're unconfirmed, and a model whose entire public case is "trust us" should be held at arm's length until the weights are open.

The two facts that matter most: no benchmarks, and open weights already slipped

Two things separate a confident launch from a proven one, and Qwen3.8-Max is missing both.

Zero published benchmarks. Alibaba released the model with no benchmark scores, no model card, no activated-parameter count, and no architecture or training detail. The "second only to Fable 5" line is the company grading its own homework. There is currently nothing to independently check — which, for a model marketed as near-frontier, is a conspicuous absence.

The open-weight date already moved. Alibaba said the full weights would be released — one widely cited timeline pointed at around July 27, 2026 — and as of writing, they're still pending. The open-weight release is the single thing that would let outside researchers confirm any of the claims above. Until it lands, this is a hosted preview you rent, not a model anyone outside Alibaba has verified.

Neither of these is a reason to ignore Qwen3.8-Max. Both are reasons to measure it yourself before you believe the pitch.

No benchmarks is exactly why a gateway matters here

When a lab publishes full benchmarks and a model card, you can at least argue about the numbers. When it publishes none, the only number that means anything is the one you generate on your workload.

That's the honest version of "top tier globally": send the same prompts to qwen3.8-max-preview, claude-*, gpt-*, gemini-* and deepseek-*, and compare latency, cost, and output side by side. Through a gateway you do that on one key, without onboarding a new provider or rewriting your client:

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_APIFLUX_KEY",
    base_url="https://apiflux.ai/v1",
)

prompt = "Build a playable browser tetris in a single HTML file."

for model in ["qwen3.8-max-preview", "claude-opus-4-8", "gpt-5.6", "deepseek-v4-pro"]:
    r = client.chat.completions.create(
        model=model,                         # confirm exact slugs at apiflux.ai/models
        messages=[{"role": "user", "content": prompt}],
    )
    print(model, "→", len(r.choices[0].message.content), "chars")

What the gateway gives you for a launch like this:

  • A/B against the incumbents on one key — replace Alibaba's missing benchmarks with your own numbers, on your own tasks.
  • Automatic failover — a preview endpoint under launch-week load will rate-limit; route around it instead of eating the errors.
  • Transparent per-token pricing — Qwen3.8-Max is currently sold inside Alibaba's Token Plan subscription rather than at a public per-token rate. The moment a per-token price exists, it shows up on the Qwen model page at the exact number, not a marked-up mystery. Until then: we won't invent one.

Availability note: qwen3.8-max-preview lands on ApiFlux as soon as the upstream is publicly stable. Watch the Models page — the exact slug and live pricing appear there the day it's routable.

Timeline

  • July 19, 2026 — Qwen3.8-Max-Preview announced at WAIC Shanghai; live as a hosted preview.
  • ~July 27, 2026 — the open-weight release one timeline pointed to; did not ship on schedule.
  • August 1, 2026 (today) — preview callable via API/Token Plan; open weights still pending; no benchmarks published.

We'll revise this post as the official version and open weights actually land — not when they're promised.

What to do if you're evaluating Qwen3.8-Max

  • Don't wait for the benchmarks — there aren't any. Run your own eval now, on the preview, behind a fallback.
  • Discount the unverifiable claims (1M context, oh-my-cli, the case studies) until there's a doc or an openable trace.
  • Watch the open-weight release, not the announcement. That's when the "second only to Fable 5" claim becomes checkable.
  • Compare per-token, not per-headline — and only once a real price exists.

The model may well be excellent. The point is that, right now, you can't know that from what Alibaba has published — you can only know it from what you measure.

FAQ

Is Qwen3.8 released, or is it a preview? As of August 1, 2026, only Qwen3.8-Max-Preview is available — a hosted preview via API. The final version and open weights have not shipped.

Is it open source? Not yet. Alibaba promised open weights (one cited timeline was ~July 27, 2026); they're still pending. Until then, the claims are single-source.

What's the context window? Alibaba has not published an official context-window spec. The "1M" figure comes from launch materials, not documentation.

Are there benchmarks? No. No benchmark scores, model card, or activated-parameter count were released. "Second only to Fable 5" is Alibaba's own assessment.

How do I call it with my existing OpenAI-style code? Through ApiFlux it's OpenAI-compatible: change the model string to qwen3.8-max-preview and keep your key. Confirm the exact slug on the Models page and the base URL in the Docs.

Sources

Enterprise AI Gateway for routing, securing, and observing every model call.

© 2026 ApiFlux. All rights reserved.