Qwen3.8 27B

Qwen3.8-27B scores 52 on the Artificial Analysis Intelligence Index at maximum reasoning effort and 61.7 on SWE-bench Pro per Qwen's own evaluation. The model leads its comparison set on agentic coding and computer-use benchmarks while trailing frontier models on Humanity's Last Exam and GPQA Diamond, reflecting a prof The Qwen3.8-27B vs Qwen3.6-27B vs Qwen3.7-Plus tables show DeepSWE 1.1 moving from 13.3 to 42.2, QwenSWEBench from 49.3 to 79.0, OSWorld-Verified from 63.9 to 84.3, Browser-use WebArena-Verified from 48.8 to 64.8, Mobile-use AndroidWorld from 70.3 to 81.9, RecreationBench from 29.8 to 47.1, SWE-MM from 25.7 to 38.6, an

Rost Glukhov's Medium piece covers the open-weight release of Qwen3.8-27B alongside the larger Qwen3.8-Max, with both weights published on Hugging Face and ModelScope. While Qwen3.8-Max is the 2.4-trillion-parameter flagship that requires cluster-scale hardware (roughly 1.2 TB even at four-bit), Qwen3.8-27B is position The article situates Qwen3.8-27B as the practical counterpart to Qwen3.8-Max, emphasizing that owning, running, and customizing the 27-billion-parameter open-weight model is what makes it 'the most important local AI release of 2026' for the developer audience, rather than the spectacle of the trillion-parameter flagsh

Alibaba's Qwen team released Qwen3.8-27B on August 14, 2026 at 15:00 UTC. The open-weight checkpoint contains 27.78 billion parameters, accepts text, images, and video, ships under Apache 2.0, and provides a native 262,144-token context window. According to Qwen's model card, the release posts sizable gains over Qwen3. Kingy.ai's launch-day review inspected the model card, configuration, license, repository metadata, official BF16 and FP8 artifacts, and third-party GGUF inventory, recalculating benchmark deltas and memory lower bounds without running inference because the Qwen Cloud managed endpoint was marked 'coming soon.' The piec

Artificial Analysis's evaluation page covers Qwen3.8 27B (xhigh), the reasoning variant from Alibaba, released in August 2026 as an open-weights model. It scores 34 on the Artificial Analysis Intelligence Index (well above the small-model median of 8), supports text, image, and video input with text output, runs a 256k On performance and cost, Qwen3.8 27B (xhigh) is described as amongst the leading models in intelligence but expensive relative to comparable open-weight peers ($0.50 per 1M input, $3.00 per 1M output, averaging $1.01 per Intelligence Index task) and notably slow at about 43 output tokens per second, while being very ve

The Local AI Zone technical analysis confirms Qwen3.8-27B was released on August 14, 2026 at 15:00 UTC by Alibaba's Tongyi Lab under Apache 2.0. It specifies 27.78 billion parameters, a hybrid attention architecture with a 3:1 ratio of Gated DeltaNet (linear) to full attention across 64 transformer blocks, a 256K nativ The analysis reports that Qwen3.8-27B achieves competitive performance with models 10–15× its size while requiring only 24GB VRAM minimum, outperforming Meta's Muse Glimmer (30B) across 8 direct benchmarks and surpassing Claude Opus 4.6 on 15 of 19 overlapping tests. Key architectural innovations include Gated DeltaNet

Ollama's library page confirms Qwen3.8 27B's availability for local deployment, listing 1.5M downloads, an 18GB model with Q4_K_M quantization (17GB), a CLIP vision projector at 461M parameters in BF16 (931MB), and Apache 2.0 licensing. The official tag is 22130167c4c2 with model architecture qwen35 and 27.3B parameter The listing highlights core improvements in coding, professional work, research, and long-horizon agentic tasks, alongside stronger autonomous planning and environment feedback handling for end-to-end task completion. It supports multimodal input including images and video, from STEM diagrams and documents to hour-scal

OpenRouter's listing confirms Qwen3.8 27B as an open-weight dense vision-language model from Qwen, suited for coding, professional workflows, research, multimodal interaction, and long-running agent tasks, with toggleable thinking mode. The model supports a 1M token context, was released on August 14, 2026, and is offe Benchmark data from the listing shows Qwen3.8 27B scoring 87.2% on GPQA Diamond via Morph and 83.8% on TAU-Bench via Alibaba Cloud International, with providers like NovitaAI and Ionstream also reporting competitive GPQA and TAU-Bench results. Uptime figures across providers range from 84.15% (Cloudflare) to 99.89% (Al