2026-08-12

Qwen3.8-Max's Open Weights Finally Landed — Text-Only, and Two Weeks Late

AIModels🌍 Asia

The Hugging Face model card for Qwen/Qwen3.8-2.4T-A95B is now up, confirming what we could only infer from the repository name in our first pass at this story: this is the promised open release of Alibaba's 2.4T-parameter flagship, previewed July 19 and benchmarked August 1 with an open-weights promise for "next week" that slipped roughly two weeks past its own deadline. The card resolves most of what we flagged as unconfirmed — architecture, context length, benchmark table — and surfaces one thing the earlier launch coverage didn't mention at all: this open checkpoint is text-only.

The architecture, precisely

Qwen3.8 is a 2.4 trillion parameter model with 95 billion activated per token (confirming the "A95B" in its own name), built on a hybrid attention design: 92 layers arranged as 23 repeated blocks, each block stacking three Gated DeltaNet (linear-attention) layers followed by one Gated Attention (full-attention) layer, every layer paired with its own mixture-of-experts FFN — 512 routed experts total, with 10 routed plus 1 shared active per token. Context runs 262,144 tokens natively, extensible to just over 1,010,000. It's trained with multi-token prediction (MTP).

That 3:1 linear-to-full attention ratio is the same architectural bet Kimi's Delta Attention popularized and Ant Group's Ling-3.0 line adopted — cheap long-context modeling from the linear-attention layers, periodic full-attention layers to keep retrieval quality from collapsing. Qwen's implementation uses Gated DeltaNet rather than KDA specifically, a related but distinct design in the same linear-attention family. Three major Chinese labs converging on structurally similar hybrids within months of each other is a stronger signal about where the field has landed on long-context efficiency than any one model's benchmark score.

The real finding: open, but not the multimodal flagship

The July 19 preview was reported — including in our own earlier coverage — as Qwen's first multimodal model above 1T parameters, with native text, image, and video input. The model card for the actual open weights says otherwise, explicitly: "Qwen3.8-2.4T-A95B is a text-only model that requires thinking mode for all interactions. Multimodal inputs are not supported, and thinking cannot be disabled." Every response begins with mandatory reasoning in <think> tags before any output — there's no non-reasoning mode to fall back to for cheap, fast answers, and no vision or video path at all.

The benchmark table on the card doesn't include the multimodal comparisons (MMMU-Pro, VideoMME, and the rest) that anchored our August 1 coverage of the API-served product, either — this card's comparison set is Claude Opus 4.8, Claude Fable 5, GPT-5.6 Sol (max), and the previous-generation Qwen3.7-Max, on coding, agentic, and general-capability benchmarks only. The overlapping numbers match our earlier reporting exactly (PaperBench 93.0, IFBench 82.8, HealthBench 60.2, SWE-bench Pro 67.7, HLE 43.6), which tells us this is the same underlying model generation — but the multimodal capability Alibaba demonstrated at WAIC appears to live in the API-served product and not in this open checkpoint. That's a real and specific gap between what's sold and what's open, in the same shape we've flagged elsewhere this month: "open" describing a narrower artifact than the flagship being marketed under the same name.

What the benchmark table actually shows

Reproducing the card's own table, since it's the first fully independent-of-secondary-sourcing look we've had:

BenchmarkOpus 4.8Fable 5GPT-5.6 Sol (max)Qwen3.7-MaxQwen3.8-Max
Terminal Bench 2.184.684.688.874.586.6
SWE-bench Pro69.280.064.660.667.7
PaperBench80.388.890.564.893.0
IFBench62.263.572.779.182.8
HealthBench52.455.354.560.2
HLE45.753.347.241.443.6
HLE w/ tools57.964.558.053.556.2
GPQA Diamond92.092.694.192.492.6
WideSearch72.981.275.281.9

The generational jump from Qwen3.7-Max is large and consistent — Terminal Bench 2.1 up 12 points, IFBench up 3.7, PaperBench up 28.2 — which is the more reliable signal in a same-family, same-methodology comparison than the cross-lab rows. Against outside models, the pattern from our August 1 coverage holds: real leadership on instruction-following and research-reproduction benchmarks (IFBench, PaperBench, WideSearch), competitive-but-not-leading on coding agents, and the weakest relative showing on HLE, the hardest general-reasoning test on the sheet. The card documents its evaluation harnesses per benchmark in detail — Claude Code for most, OpenCode for the Qwen-series SkillsBench run — which is the kind of methodology disclosure we credited Ling-3.0-tiny's card for and is genuinely useful here for the same reason.

Still unconfirmed

The card doesn't name a license anywhere in the sections available to us — Qwen 3.5 and 3.6 shipped Apache 2.0, and nothing here contradicts that, but nothing confirms it either. Whether Qwen3.8-27B, the smaller sibling promised in the same August 1 announcement, shipped in the same window remains unverified. And whether Alibaba plans to open-weight a multimodal variant of Qwen3.8-Max at all, or is keeping that capability API-only indefinitely, is now the more interesting open question than the one we started with.

What to expect next

  • Watch for the license text. Apache 2.0 on a genuine 2.4T model, even sparse at 95B active, would still be the most permissive large-scale release of the year.
  • Watch whether a multimodal open checkpoint ever ships. The gap between the marketed capability and the open artifact is the real story here — if Alibaba never open-weights the vision/video version, "open" and "flagship" have quietly become two different products under one name.
  • Watch for independent reproduction of the coding and agentic numbers, now that the weights are actually downloadable and auditable rather than API-gated.
  • Watch the deployment reality, as we flagged before: 2.4T total parameters is a serious infrastructure commitment even at 95B active, and "open" doesn't mean "runnable by most people who'd want to."

References: Hugging Face — Qwen/Qwen3.8-2.4T-A95B model card · MarkTechPost — the July 19 preview · related coverage: Qwen3.8-Max previews with zero benchmarks · Qwen3.8-Max showed its work · Ling-3.0-tiny · Levels of Open · Open weights you cannot run · Frontier Arcade: trends & predictions