On August 12, Alibaba published the open weights of Qwen3.8-Max on Hugging Face: two official repositories, Qwen/Qwen3.8-2.4T-A95B and an FP8 quantized variant, a fine-grained Mixture-of-Experts with 2.4 trillion total parameters and 95 billion active. It is a post-trained checkpoint, with the model card explicitly targeting vLLM, SGLang, and TokenSpeed deployment paths. NVIDIA confirmed the release the same day in a deployment engineering blog. This is the first time the Qwen family has handed out a Max-tier flagship — but once you download it, you will find it is not the model you tried in the API.
What landed is not the flagship, but a text-only subset
Compared against the hosted Qwen3.8-Max, the open-weights release is missing three things: vision input (available on the API, absent here — the checkpoint is text-only), the up-to-1M-token context window (the native window is documented well below that), and optional thinking mode (the open version forces thinking on for all interactions). Tool use and the Qoder agent surface are also not part of the base checkpoint.
The gap immediately lit up the community. A Hugging Face discussion thread opened within hours of the drop: commenters argued that stripping vision "removes half its core value," others compared the same-name-different-capabilities release pattern to game-industry DLC paywalling, with at least one saying it cost Alibaba "all good will" built up across earlier open Qwen generations.
Hardware requirements keep most people out anyway. The full BF16 checkpoint weighs around 4.89 TB; Unsloth GGUF quants span roughly 397 GB at 1-bit to about 1.31 TB at 4-bit — even pushed to the floor, that is multi-GPU workstation or heavy NVMe-offload territory. NVIDIA reference deployment is a GB300 NVL72 rack (72 Blackwell Ultra GPUs), serving the FP8 checkpoint at over 4,000 tokens/sec per GPU. For single-card users, this release has essentially nothing to do with local deployment.
The license is the real signal
More telling than the feature cuts is the license change. Earlier Qwen generations shipped under permissive Apache 2.0 terms; this time the HF repo carries a bespoke qwen3.8-max license, and reporting around the launch describes a revenue-sharing requirement for large commercial users, with the exact threshold and percentage still being finalized. Hugging Face Summer 2026 State of Open Models report, published August 14, corroborates the direction: the very top of the frontier is beginning to shift toward clearer monetization, with Kimi K3 and Qwen 3.8 Max recently adding non-commercial restrictions and revenue-share requirements.
That explains the "why" behind the cuts. When open weights shift from a customer-acquisition channel into an asset that must monetize directly, flagship capabilities have to stay on the API side. Alibaba still delivered the first Max-tier open release in Qwen history — but in the definition of "open," this drop reads more like a research-infrastructure showcase than a full handoff to the community.
The 27B shipped, but the real question remains
On August 15, the promised single-GPU companion Qwen3.8-27B also went live (it hit #1 on Hacker News the day it shipped), under Apache 2.0, with no feature stripping this time. That gives the "missing half" a landing spot — self-hosters now have a genuinely accessible option.
But for the industry at large, the underlying question is unchanged: when top-tier openness turns from full delivery into tiered delivery, the open-source community gets the weights without getting capabilities on par with the closed version. What you tested in Qoder is the full model; what you download locally is the stripped one. From now on, evaluating any "open flagship" starts with one question: which half is actually open? Reference: explainx.ai analysis