Contents
Two weights, one generation
Alibaba opened the weights for both members of its Qwen3.8 generation this week: a 27-billion-parameter dense model that is natively multimodal, plus a 2.4-trillion-parameter mixture-of-experts flagship that activates about 95 billion parameters during inference. ² The generation was announced on August 3, 2026, when Qwen3.8-Max went live with a priced API and Alibaba promised open weights ‘within about a week.’ ³ Both models’ weights appeared on Hugging Face on August 14, and the 27B repository also landed on ModelScope that evening Beijing time, closing the open-date question. ²³ The 27B is the dense, self-hostable sibling positioned against the Max flagship and the successor to the widely used Qwen3.6-27B. ³ Qwen frames the drop as its first open release of ‘Max-class’ capability — a tier that previously lived exclusively behind a commercial API. ² It has drawn a fast crowd already: the 27B repository recorded over 267,000 downloads shortly after release. ¹
A hybrid attention stack with adjustable reasoning
Both models run on a refreshed architecture that interleaves fast linear layers with full self-attention blocks — a repeating pattern of three Gated DeltaNet layers followed by one full attention layer. ² The 27B variant stacks 64 transformer layers at a hidden dimension of 5120, while the 2.4T model pushes to 92 layers and a hidden dimension of 8192. ² On Hugging Face the 27B is registered as an image-text-to-text model with roughly 27.8 billion parameters in BF16. ¹ It is natively multimodal across text, images, video, diagrams, and documents, with a native 262,000-token context that extends to 1 million tokens via YaRN, plus a configurable reasoning_effort mode that adjusts thinking depth to save compute. ³ The MoE has 512 routed experts, activating 10 plus a shared expert per token to hold effective compute near 95B. ² On the 27B, thinking can be disabled for routine prompts; the large MoE variant always reasons. ²
Vendor-reported evaluations
Alibaba’s own evals show the 27B clearing its predecessor and even the larger Qwen3.7-Plus: SWE-bench Pro 61.7 versus 53.5 for Qwen3.6-27B, agentic terminal coding 73.0 versus 63.4, and JobBench 33.4 versus 21.8. ³ These are vendor-reported numbers; no independent benchmark has scored the 27B so far, so they amount to a claim the community will now need to reproduce. ³ The model card also lists 42.2 on DeepSwe, 89.2 on GPQA Diamond, 30.8 on HLE, and 57.4 on ClawEval-MM, with the SWE-bench Pro figure evaluated using a Claude Code harness. ¹ The larger MoE model recorded 67.7 on SWE-bench Pro, 56.6 on DeepSWE 1.1, and a PaperBench 93.0 — the highest across the benchmarks tested — though it still trails on the hardest coding tasks. ²
Licensing and day-one ecosystem
Licensing is where the two releases split. The 27B ships under Apache 2.0, a permissive license with no user-count cap and no commercial-discussion trigger, so the earlier worry that a Qwen release would fall under the Tongyi Qianwen license — with its 100-million-MAU clause — did not materialize for the dense model. ³ The 2.4T flagship instead sits under a custom license, reflecting its Max-tier status and giving the generation a two-tier licensing structure. ² That structure makes the 27B the low-obstacle path into products and agents. ³ At launch, SGLang, vLLM, and TokenSpeed supported inference out of the gate, and quantizers were quick: 17 community variants per model are already up across llama.cpp, Ollama, LM Studio, and Jan. ² The 27B also appeared on a commercial API marketplace within hours of the drop — a visible listing that shows the model is being offered through managed APIs, not just as a download. ³