--- 格式版本: 2 标题: "Qwen3.8-27B Practical Guide: Control Reasoning Depth and Extend Context to 1M Tokens" 原文链接: "https://www.alibabacloud.com/blog/qwen3-8-27b-practical-guide-control-reasoning-depth-and-extend-context-to-1m-tokens_603509" 发布日期: "2026-08-28" 发布时间校准状态: "found" 发布时间需复核: "否" 发布时间来源: "rule:configured_publication_date_rule" 发布时间证据: "alibaba-cloud-news-publication-date html:original: Alibaba Cloud Community August 28, 2026" 发布时间校准原因: "信源发布日期识别规则直接确认发布时间" 发布时间校准置信度: "high" 发布时间候选数量: 2 发布时间严格候选数量: 2 发布时间原页读取状态: "source template page reused from URL open" 发布时间未找到原因: "" 发布时间校准时间: "2026-08-31T02:22:36+08:00" 发布时间仲裁状态: "skipped" 发布时间仲裁尝试次数: 0 发布时间仲裁耗时毫秒: 0 发现时间: "2026-08-31T02:21:35+08:00" 入库时间: "2026-08-30T18:23:07.677Z" 来源平台: "固定入口" 搜索渠道: "fixed_url" 搜索词: "https://www.alibabacloud.com/blog" 匹配关键词: - "deployment" - "performance" - "latency" - "throughput" - "AI" 相关厂家: - "阿里" - "OpenAI" 相关专家: [] 内容类型: "网页" 抓取工具: "Free Fetch + Defuddle" 清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取" 原始附件: [] 图片摘要: - "✓ ./assets/img-06208895.jpg | other | " AI优质: "否" AI打分: 36 AI分档: "非优质" AI质检状态: "不通过" AI打分理由: "正文主线是Qwen3.8-27B的推理深度控制与YaRN长上下文配置教程，并非超节点或机架级AI基础设施。来源为阿里云官方博客入口并追溯至QwenDevs，但发布日期缺失。相较知识库中仅提及该模型的材料，新增了262K原生上下文、扩展至1M及vLLM/SGLang/TokenSpeed配置参数；这些属于模型Serving使用方法，未提供机架拓扑、互连、功耗、RAS、生产级基准、客户部署或商业里程碑。命中“教程与运维选型”及“应用与模型效率”硬否决项，当前页面不适合作为超节点业务情报源。" AI质检模型: "gpt-5.6-sol" AI质检时间: "2026-08-31T02:24:17+08:00" AI主题相关性: 2 AI来源权威性: 11 AI新颖性: 12 AI技术细节: 3 AI商业部署信号: 0 AI完整性: 8 AI评分提示词版本: "v17-精简生产版" AI评分提示词SHA256: "48fb9777f386026761b4873eaff30807694fb11e9b352d7c69bf2dfde750cc7d" AI评分知识库版本: "knowledge_base_v1-20260819+runtime.56" AI评分知识库SHA256: "4fc2bc9ef1ae3cccce42c5a4bdc8532ff011e9c8cef432980972e4d4a1d22131" AI评分知识库检索词: "[\"阿里\",\"https://www.alibabacloud.com/blog\",\"NPU\",\"Qwen3.8-27B\",\"assets/img-06208895.jpg\",\"Qwen3.5\",\"Qwen3.8\",\"API\",\"Qwen/Qwen3.8-27B\",\"SGLang\",\"VLLM_ALLOW_LONG_MAX_MODEL_LEN\",\"SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN\"]" AI评分知识库命中: "[{\"id\":\"runtime-5042ddb06fd844309130b6e9\",\"title\":\"Serve Qwen3.8-2.4T-A95B, a 2.4T-Parameter Model, with Configurable Reasoning on NVIDIA GB300 NVL72\",\"sourceType\":\"ai_excellent_article\",\"time\":\"2026-08-12\",\"matchedTerms\":[\"Qwen3.8\",\"SGLang\"],\"rank\":-15.878812008663186},{\"id\":\"runtime-e2c7461df9a3b0d18402efe7\",\"title\":\"AliViews: Eddie Wu on Alibaba's Q1 Earnings\",\"sourceType\":\"ai_excellent_article\",\"time\":\"2026-08-21\",\"matchedTerms\":[\"Qwen3.8-27B\",\"Qwen3.8\",\"API\"],\"rank\":-10.090192342622139},{\"id\":\"july-correct-0001\",\"title\":\"全球首颗2nm GPU来了！苏姿丰甩出“最强AI机架”，CPU性能干翻英伟达 - 智东西\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"阿里\",\"NPU\",\"SGLang\"],\"rank\":-9.23940547478323},{\"id\":\"july-correct-0088\",\"title\":\"StrataCL: Fabric-Native Communication Library for Production Supernodes - 智源社区论文\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"NPU\"],\"rank\":-6.195219884264695},{\"id\":\"historical-jun-025\",\"title\":\"英伟达、谷歌与国产超节点的三种网络选择\",\"sourceType\":\"curated_item\",\"time\":\"2026-06\",\"matchedTerms\":[\"NPU\"],\"rank\":-5.819082117911849}]" AI摘要: "阿里云发布Qwen3.8-27B实用指南，该模型支持通过reasoning_effort按请求调节推理深度，并原生支持262K token上下文，可用YaRN扩展到100万token。" AI摘要模型: "ali-deepseek-v4-flash" AI摘要时间: "2026-08-30T23:29:23.539Z" 采集批次: "2026年8月31日2点04分09秒" 采集批次ID: "20260831-020409-922" 去重键: "https://www.alibabacloud.com/blog/qwen3-8-27b-practical-guide-control-reasoning-depth-and-extend-context-to-1m-tokens_603509" --- ![1](./assets/img-06208895.jpg) Qwen3.8-27B is here — a compact, deployment-friendly dense model built on the Qwen3.5 architecture, with native vision-language understanding and flexible thinking control. Two developer-facing features are worth your attention right away: 1. **reasoning\_effort** — dial reasoning depth up or down, per request. 2. **Ultra-long context** — 262K tokens natively, extensible to 1M with YaRN. Here's how to use both. ## 1\. reasoning\_effort: Control Reasoning Depth Per Request Qwen3.8 thinks by default before answering. With official support for reasoning\_effort, you can now tune how hard it thinks — balancing accuracy, speed, and cost: - xhigh (default): for complex tasks that require thorough analysis - medium: for balancing accuracy and speed - low: for efficient reasoning that prioritizes speed and cost A minimal example via the OpenAI-compatible Chat Completions API: ``` from openai import OpenAI # Configured by environment variables client = OpenAI() messages = [{"role": "user", "content": "Write a Python function to merge two sorted linked lists."}] completion = client.chat.completions.create( model="Qwen/Qwen3.8-27B", messages=messages, extra_body={ "chat_template_kwargs": { "enable_thinking": True, # on by default "preserve_thinking": True, # on by default }, }, reasoning_effort="xhigh", # supported levels: xhigh, medium, low ) ``` One practical tip: in multi-turn agentic tasks, lower effort doesn't always mean faster end-to-end. Faster per-turn responses may come with more failures and retries, which can increase total latency and token consumption. Pick the effort level that fits the task. ## 2\. Processing Ultra-Long Contexts: 262K Native, Up to 1M with YaRN Qwen3.8-27B natively supports context lengths of up to 262,144 tokens. When your total length (input + output) exceeds that, enable RoPE scaling — YaRN is supported by vLLM, SGLang, and TokenSpeed. The quickest approach is to pass the configuration as engine startup flags. For vLLM, use ``` VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve ... --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --max-model-len 1000000 ``` For SGLang, use ``` SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python -m sglang.launch_server ... --json-model-override-args '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --context-length 1000000 ``` For TokenSpeed, use ``` TOKENSPEED_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 tokenspeed serve ... --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --max-model-len 1000000 ``` Or modify rope\_parameters inside text\_config in the model's config.json: ``` { "mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144 } ``` Two important notes: - Open-source frameworks implement **static YaRN** — the scaling factor stays constant regardless of input length, which can hurt performance on shorter texts. Only enable YaRN when you actually need long contexts. - Tune factor to your typical context length. If your workloads usually sit around 524,288 tokens, set factor to 2.0 instead of 4.0. ## Deploy It Your Way Qwen3.8-27B is compatible with vLLM, SGLang, TokenSpeed, Unsloth, and more. For production workloads or high-throughput scenarios, see the deployment guides maintained by each framework: - **vLLM**: [https://recipes.vllm.ai/Qwen/Qwen3.8-27B](https://recipes.vllm.ai/Qwen/Qwen3.8-27B) - **SGLang**: [https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8-27B](https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8-27B) - **TokenSpeed**: [https://lightseek.org/tokenspeed/recipes/models#qwen3-8](https://lightseek.org/tokenspeed/recipes/models#qwen3-8) - **Unsloth**: [https://unsloth.ai/docs/models/qwen3.8](https://unsloth.ai/docs/models/qwen3.8) Happy building. --- *[Source](https://x.com/QwenDevs/status/2088289608031510811)* ![1](./assets/img-06208895.jpg)