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, leading its comparison set on agentic coding and computer-use benchmarks while trailing frontier models on Humanity's Last Exam and GPQA Diamond. The page labels which hea The excerpt provides Qwen3.8-27B benchmark figures from the model card against Qwen3.6-27B and Qwen3.7-Plus, including Terminal-Bench 2.1 at 73.0, SWE-bench Pro at 61.7, NL2Repo-Bench at 42.3, DeepSWE 1.1 at 42.2, QwenSWEBench at 79.0, CoWorkBench at 70.7, JobBench at 33.4, IFBench at 79.5, GPQA Diamond at 89.2, Humani

Vercel's open-weight models guide, published about three days before the current date, frames the strategic context in which a dense model like Qwen3.8 27B is offered through AI Gateway. It notes that open-weight models accounted for 36% of AI Gateway token volume in July 2026, up from 11% in April, at roughly one-seve The guide distinguishes open-weight from open source, emphasizing that weights ship without the training code, data, or process, and that the license governing production use often differs from the family's headline license across generations and model sizes. It also warns that the model string is the only portable par

Intel's OpenVINO toolkit publication announced Day-0 support for Qwen3.8-27B on Intel Agentic PCs, framing the model as a relatively compact, deployment-friendly dense multimodal vision-language model with image and video understanding, switchable Thinking Control, and design for complex multi-step agent tasks. The art The piece positions Qwen3.8 as targeting leading performance across coding, professional workloads, scientific research, and long-horizon agent tasks, with Qwen3.8-27B extending those advances into a compact dense checkpoint. It highlights stronger autonomous planning and environmental feedback adaptation for agentic t

Qwen3.8-27B was released by Alibaba's Qwen team on August 14, 2026 as a 27-billion-parameter dense multimodal model under Apache 2.0, with native multimodal capabilities and 262K-token context extensible to 1M tokens via YaRN. The excerpt positions the 27B variant as more practically relevant than its 2.4T flagship cou The page lists specifications of 27.78 billion parameters, a dense architecture, native 262K context, up to 1M tokens via YaRN, native vision for image and video, reasoning/thinking mode, and coding/agentic capabilities, and reports large gains over Qwen3.6-27B on coding benchmarks, with DeepSWE moving from 13.3 to 42.

Qwen3.8 27B is the compact, open-weight release in Alibaba's Qwen3.8 generation, published on Hugging Face on August 14, 2026 under Apache 2.0. It is a 27-billion-parameter dense model that reads images and video natively, holds 262K tokens of context extensible to 1M with YaRN, and is positioned as the deployment-frie The excerpt reports Qwen's own launch-table results, including 61.7% on SWE-bench Pro versus a listed Claude Opus 4.6 Max of 53.4%, 73.0 on Terminal-Bench 2.1, and 84.3% on OSWorld-Verified computer use. It also cites a LuminaBench composite index shift for the 27B model from 42.2 (rank 67) on Qwen3.6 27B to 67.3 (rank

Qwen3.8-27B was released on August 14, 2026 at 15:00 UTC by the Qwen Team at Alibaba Cloud's Tongyi Lab, shipping under Apache 2.0 with the model card on Hugging Face. The excerpt characterizes it as a 27.78-billion-parameter dense multimodal language model designed for local deployment, with a stated 24GB VRAM minimum The article details a hybrid attention architecture built on the Qwen3.5 foundation: 64 transformer blocks with hidden dimension 5,120, a 3:1 ratio of Gated DeltaNet linear-attention layers to gated full-attention layers (48 vs 16), 24 query heads with 4 KV heads, head dimension 256, vocabulary 248,320, native 262,144-

Alibaba's Qwen team released Qwen3.8-27B at 15:00 UTC on August 14, 2026, and the excerpt describes it as a strong candidate for the best dense, locally deployable multimodal model around 30 billion parameters. The exact checkpoint contains 27.78 billion parameters, accepts text, images and video, ships under Apache 2. The piece cites official model-card gains over Qwen3.6-27B in agentic coding, computer use, and vision-language work, with Terminal-Bench 2.1 rising from 63.4 to 73.0, DeepSWE 1.1 from 13.3 to 42.2, OSWorld-Verified from 63.9 to 84.3, and SWE-MM from 25.7 to 38.6. It also flags caveats: every launch score comes from Qw

Artificial Analysis published an independent evaluation of Qwen3.8 27B in its "xhigh" reasoning configuration, rating it 34 on its Intelligence Index against a class median of 8. The model is listed as an open-weights release from Alibaba dated August 2026, with 27 billion total parameters, a 256k token context window, Pricing and throughput figures are provided: $0.50 per 1M input tokens and $3.00 per 1M output tokens (with a 90% cache discount, $0.82 effective), versus medians of $0.05 and $0.15 respectively, and a measured output speed of 42.6 tokens per second. The page also flags the model as notably verbose and expensive relati