Qwen3.8 27B

Regolo.ai provides a head-to-head benchmark comparison of Qwen3.8-27B against Claude Opus 4.6 Max across 24 benchmarks from the official Alibaba model card. Qwen3.8-27B wins 16, with largest gains in agentic coding (SWE-bench Pro 61.7 vs 53.4), instruction following (IFBench 79.5 vs 62.5), and computer use (OSWorld-Verified 84.3 vs 72.7). It loses on knowledge-heavy reasoning such as Humanity's Last Exam (30.8 vs 40.0) and GPQA Diamond. The article confirms Qwen3.8-27B is a dense 27-billion-parameter open-weight vision-language model with 28 billion parameters including the vision encoder, running on a single 24GB GPU at 4-bit quantization. It also notes that against Muse Glimmer-30B, Qwen3.8-27B wins every overlapping benchmark row. All figures are explicitly vendor-reported by Alibaba as of August 14, 2026, with no independent reproduction existing at launch.

Qwen3.8 27B is an open-weight dense model from the Qwen team highlighted in The New Stack as delivering near Opus 4.6-class performance while running on a laptop. The piece frames the release as a local-inference milestone, giving developers a strong coding-capable model without cloud dependence. This positions 27B as a practical desktop alternative to larger hosted systems. The article confirms Qwen3.8 27B targets local deployment and professional workflows, reinforcing its open-weight availability for the developer community. It underlines Alibaba's Qwen team's continued emphasis on efficiency at the 27B parameter scale. This makes the model relevant for teams prioritizing on-device inference and cost control.

Qwen3.8-27B is a dense 27.78-billion-parameter multimodal model released by Alibaba's Qwen team on August 14, 2026, under Apache 2.0. The Kingy.ai launch-day review confirms a native 262,144-token context window, with text, image, and video input support. Vendor-reported benchmark deltas versus Qwen3.6-27B include Terminal-Bench 2.1 rising from 63.4 to 73.0, DeepSWE 1.1 from 13.3 to 42.2, and OSWorld-Verified from 63.9 to 84.3. The review flags that all launch scores are self-reported by Qwen and that several benchmarks are in-house or modified, with the SWE-bench Pro Opus comparison imported rather than rerun. It also warns that local-hardware claims often ignore KV cache memory and that the 1M-token context requires YaRN scaling rather than the native window. Official BF16 and FP8 artifacts plus third-party GGUF files were inspected, but no inference was run at launch due to the Qwen Cloud endpoint being marked "coming soon."

OpenRouter's listing confirms Qwen3.8 27B is 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. It supports a 262K-token context and was released on August 14, 2026. These attributes establish the model's core capability profile. The page surfaces provider-by-provider routing metrics including latency (0.41s P50), throughput (34 tok/s), and benchmark percentages on GPQA Diamond, TAU-Bench, and Darkbloom across hosts such as ModelRun, Cloudflare, DekaLLM, and Parasail. While useful for deployment decisions, these figures describe serving-provider performance rather than intrinsic model characteristics.

Local AI Zone provides a comprehensive technical analysis of Qwen3.8-27B, confirming it as a 27.78-billion-parameter dense multimodal model from Alibaba's Tongyi Lab released August 14, 2026 under Apache 2.0. The model uses a hybrid attention architecture with 48 Gated DeltaNet linear-attention layers and 16 full-attention layers, totaling 64 transformer blocks with a 5,120 hidden dimension. Native context is 262,144 tokens, extendable to 1 million via YaRN. The analysis reports Qwen3.8-27B achieves competitive performance against models 10-15x its size, running on as little as 24GB VRAM. It outperforms Meta's Muse Glimmer (30B) across all 8 direct comparison benchmarks and surpasses Claude Opus 4.6 on 15 of 19 overlapping tests. Key architectural features include multi-token prediction for speculative decoding, 24 query heads with 4 KV heads, and a fused attention output gate, signaling a new efficiency frontier for local AI deployment.