Qwen releases Qwen3.8 27B multimodal model under Apache 2.0

• Qwen3.8-27B puts a 27B dense vision-language model under Apache 2.0, with 262K native context and adjustable reasoning, as WesRoth's release rundown reports.
• Independent 128K testing found medium reasoning reached 92.8% AUC, while xhigh spent more than 12 times as many thinking tokens and scored lower, per DillonUzar's Context Arena test.
• Long-context accuracy still drops sharply at the endpoint: medium, low, and xhigh landed at 78.9%, 81.8%, and 74.3% pointwise accuracy at 128K in DillonUzar's per-mode results.
• The local rollout is already broad: Alibaba_Qwen's Ollama announcement put the model on Ollama, while UnslothAI's Desktop post added local running and training.

An NVIDIA Developer Forum deployment note reports that an early NVFP4 package silently truncated prompts at 2,048 tokens before a fixed upload. Qwen's official model card lists a managed Qwen Cloud version with 1M context and built-in tools as coming soon. In his local review, simonw ran the same 17GB Q4_K_M quantization on an M5 Max MacBook Pro and a DGX Spark.

• The Qwen3.8-27B weights ship under Apache 2.0, according to Qwen's official model card, and WesRoth's release rundown identifies the checkpoint as a 27B dense model.
• Native multimodal input covers text, images, and video, according to WesRoth's release rundown.
• The native context window is 262K tokens, with extension to roughly 1M tokens via YaRN in WesRoth's release rundown.
• Thinking is adjustable by effort level, and Qwen says reasoning can persist across multi-turn agent sessions in WesRoth's release rundown.
• Qwen positioned the checkpoint for laptop use in Alibaba_Qwen's launch post.

First-party

• Terminal-Bench 2.1: 63.4% → 73.0%, +9.6 points, in WesRoth's Qwen benchmark chart.
• SWE-bench Pro: 53.5% → 61.7%, +8.2 points, in WesRoth's Qwen benchmark chart.
• DeepSWE 1.1: 13.3% → 42.2%, +28.9 points, in WesRoth's Qwen benchmark chart.
• OSWorld-Verified: 63.9% → 84.3%, +20.4 points, in WesRoth's Qwen benchmark chart.

Third-party evaluators

• GDM-MRCRv2 AUC at 128K: Qwen3.6-27B with thinking 74.5% → Qwen3.8-27B with medium thinking 92.8%, +18.3 points, per DillonUzar's Context Arena test.

Customer-reported

• Private benchmark: GLM 5.2's undisclosed score → Qwen3.8-27B's undisclosed score, +4%, according to onusoz's early test, whose author also questioned the benchmark.

• Medium reasoning at 64K → 128K: 97.9% → 78.9% pointwise accuracy, -19.0 points, per DillonUzar's per-mode results.
• Low reasoning at 64K → 128K: 93.3% → 81.8% pointwise accuracy, -11.5 points, per DillonUzar's per-mode results.
• Xhigh reasoning at 64K → 128K: 87.4% → 74.3% pointwise accuracy, -13.1 points, per DillonUzar's per-mode results.
• Medium → xhigh at 128K AUC: 92.8% → 86.6%, -6.2 points, while estimated thinking tokens per sample rose from 568 to 7,019 in DillonUzar's Context Arena test.
• Thinking off → medium thinking at 128K AUC: 43.1% → 92.8%, +49.7 points, in DillonUzar's Context Arena test.
• The independent run stopped at 128K because available providers exposed roughly 256K context, leaving insufficient room for outputs at the longer bin, according to DillonUzar's follow-up.

• The checkpoint is dense rather than MoE, with native vision-language processing for image and video input, as WesRoth's release rundown describes.
• Qwen describes 262K tokens as the native window and roughly 1M as the YaRN-extended window in WesRoth's release rundown.
• The reasoning controls are low, medium, and xhigh, and Qwen says preserved reasoning lets an agent carry prior reasoning through a multi-turn session in WesRoth's release rundown.

• simonw configured Pi to use Qwen3.8-27B and had it build a script that converts Pi's own JSONL transcripts to Markdown, according to simonw's Pi experiment.
• On an M4 Max with 64GB, tomgreenwald measured Qwen at a 32-second prefill and about 15 tokens per second on the same repository where a 30B-A3B Nemotron MoE took six seconds and about 70 tokens per second in tomgreenwald's hardware comparison.
• anderslie reported 10 to 20 tokens per second for a local M4 Max run, while UnslothAI's volcano simulation showed a 4-bit GGUF generating an interactive physics-heavy volcano simulation.

• Ollama listed Qwen3.8-27B as available in Alibaba_Qwen's Ollama announcement.
• Unsloth Desktop added local inference and training support in UnslothAI's Desktop post, alongside the project's local-model tooling.
• ggerganov's llama-server recipe pairs Q4_K_M with a Q4_0 draft model and enables MTP speculation, preserved reasoning, and agent mode in ggerganov's serving recipe.
• Magnitude offered a one-click local run in the setup behind anderslie's Magnitude report.