Qwen's 27B Can See. Its 2.4T Flagship Cannot.

Alibaba's first Max-class open weights arrive with no vision stack, thinking locked on, and a license that names its own products.

On August 3, Alibaba launched Qwen3.8-Max as an API and promised something no Max-class Qwen had ever done: open weights, the following week. The promise landed in two cartons. Around August 12 to 13, Qwen3.8-2.4T-A95B went live on Hugging Face, 2.4 trillion parameters of sparse mixture-of-experts, the base model behind the Max API and, after Kimi K3's 2.8T, the second-largest open-weight mixture-of-experts ever published. On August 14, Qwen3.8-27B followed, a dense model small enough for a single workstation.

Open the cartons, though, and the contents run backwards. The 27 billion parameter model ships with a native vision encoder that reads images and hour-scale videos, a thinking mode you can tune, and a clean Apache 2.0 license. The 2.4 trillion parameter flagship ships with none of that: text in, text out, reasoning permanently switched on, and a custom license titled the Qwen3.8-Max License. The small box got the eyes and the paperwork. The big box got a customs form.

This issue is the manifest: what is in each carton, what was withheld, what the license actually says, and what the download counters already show about which release mattered.

The 2.4T checkpoint is the first full look at how Alibaba builds a Max-class model, and the model card publishes the layout down to the head counts. The network is 92 layers arranged as a repeating four-layer block: three layers of Gated DeltaNet, a linear-attention design whose cost grows steadily with sequence length, then one layer of conventional Gated Attention. Every layer feeds a mixture-of-experts stage with 512 experts, of which 11 fire per token (10 routed plus 1 shared). That routing is why a 2,446 billion parameter model activates 95 billion per token, roughly 4 percent of itself.

The 27B is the same architectural family scaled down and made dense: 64 layers interleaving the same Gated DeltaNet and Gated Attention blocks, hidden size 5,120, the same 262,144-token native context extensible toward a million with RoPE scaling, plus the piece the flagship checkpoint lacks, a native vision tower. Hugging Face lists it at 27.8 billion parameters with the vision encoder included: 55.57 gigabytes in BF16, about 17 gigabytes at 4-bit per Unsloth, which fits a single 24-gigabyte card.

The Max API is multimodal: images, video, optional thinking, a million tokens of context by default, built-in tools. The model card for the open 2.4T checkpoint states, in its own words, that the hosted Qwen3.8-Max is the official version of this same base model with more features. Then it lists what the downloadable version does without. The weights are text in and text out. Thinking cannot be disabled: every single response begins with a reasoning block, and the card recommends budgeting up to 262,144 tokens for the reasoning alone plus 131,072 for the answer. And the million-token default stays in the cloud; the weights give you 262,144 natively and a documented extension path.

The community read the table the same way. The top pinned discussion thread on the repo calls the weights stripped of the Max features, and points out that the license metadata tag on the page literally reads qwen3.8-max. The same author later updated with a more balanced take: the model remains usable at no cost through chat.qwen.ai, which Kimi K3 never matched, and publishing a 2.4 trillion parameter checkpoint at all is a milestone. Both readings are correct. That is what makes the release interesting.

Alibaba gave no reason for the subtractions. The neutral reading is serving pragmatism: a vision tower on a 2.4T checkpoint adds size and integration work. The commercial reading is sharper: every withheld feature (vision, the thinking toggle, the million-token default, managed tools) is a line item the Qwen Cloud API charges for. The delta between the carton and the cloud is Alibaba's own map of where it thinks the revenue lives.

The Qwen3.8-Max License is one page, and most of it reads like MIT: use, copy, modify, merge, publish, distribute, sublicense, host, fine-tune, build derivatives, commercially or otherwise. Then come two conditions, and the second one is the first of its kind that we know of.

Moonshot's Kimi K3 license, which we covered on July 28, drew a similar line at 20 million dollars, but scoped it to model-as-a-service revenue: hosting the weights for others. Qwen raises the threshold to 50 million dollars and widens the category to reach products. A coding assistant or an office-productivity assistant built on these weights, above that revenue, must negotiate with the vendor whose Qoder and QwenWork occupy exactly those two markets. The license does not forbid competing. It prices competing, and it names the categories by pointing at Alibaba's own product lineup.

The 27B carries none of this. It is plain Apache 2.0, the same license every recent open Qwen has shipped under. The strategic clause sits solely on the artifact that carries the frontier capability.

Open weights and runnable weights are different artifacts. The 2.4T checkpoint is 4.89 terabytes of BF16 safetensors across 213 shards, with an official FP8 build at 2.5 terabytes. NVIDIA published a deployment walkthrough for serving it on a GB300 NVL72 rack, which tells you the intended audience. For everyone else, Together AI and Fireworks already serve the open checkpoint as an API, which preserves the practical benefits of open weights (no vendor lock on the artifact, reproducible evals, fine-tune rights) without the rack.

Two days after release, the 27B has roughly 34 times the downloads of the flagship and went to number one on Hacker News. The 2.4T carton is the headline; the 27B carton is the adoption. Which was probably the design: the artifact people can run is the clean Apache one, and the artifact that competes with frontier APIs is the one wearing the customs form.

The model card ships a benchmark table that puts Qwen3.8-Max at 86.6 on Terminal-Bench 2.1, 67.7 on SWE-bench Pro, 93.0 on PaperBench and 82.8 on IFBench, against columns for Claude Opus 4.8, Claude Fable 5 and GPT-5.6 Sol. The vendor numbers are strong. The footnotes are where the reading gets interesting, and to Qwen's credit, the footnotes are unusually complete.

Start with the harness. Nearly every agentic row was evaluated by running Qwen's own model inside Claude Code, Anthropic's coding agent, and the DeepSWE footnote states outright that Qwen3.8-Max performs best on Claude Code among the harnesses tried. Meanwhile the comparison columns mix sources: the Terminal-Bench footnote says competitor scores are the best published result across harnesses (Terminus 2 for the Claude models via Artificial Analysis, Codex for GPT-5.6 Sol), while Qwen's own 86.6 ran with a 5-hour timeout and a 131,072-token output cap. On SkillsBench, each model family ran in a different harness, all measured by Qwen's own testing. And a footnote under the whole table notes that the Fable 5 results may involve fallbacks.

This matters because we have already watched the harness move this exact number. When Vals ran the API model on Terminal-Bench with the benchmark's original timeouts preserved, it scored 67.4, a 19.2-point gap against the same model's self-reported 86.6, and placed it 10th of 43 on their index at 66.1, second among open-weight models. On the Frontend Code Arena it debuted 4th at 1,668 Elo, per the Latent Space roundup. None of that makes the model weak. It makes every agentic score harness-conditional, and this card, unusually, admits it in print.