{"slug": "wildclawbench-ranks-alibaba-s-qwen3-8-27b-15th-on-agent-tasks", "title": "WildClawBench ranks Alibaba's Qwen3.8-27B 15th on agent tasks", "summary": "Alibaba's Qwen3.8-27B ranked 15th among 33 models in WildClawBench, scoring 48.0% across 60 agent tasks run through the OpenClaw harness, according to the third-party benchmark. The 27-billion-parameter model trailed larger or hosted entries from OpenAI, Anthropic, xAI, Moonshot AI, and Alibaba's Qwen3.8-Max, but beat its Qwen3.6-27B predecessor. WildClawBench recorded a 516-minute completion time but listed cost as unavailable, leaving self-hosting economics outside the comparison.", "body_md": "Alibaba's Qwen3.8-27B finished 15th among 33 models in [WildClawBench](https://internlm.github.io/WildClawBench/?ref=runtimewire), scoring 48.0% across 60 agent tasks run through the OpenClaw harness. The third-party result gives developers a more restrained measure of the model than Alibaba's own coding and computer-use evaluations.\n\nThe 27-billion-parameter model finished behind larger or hosted entries from OpenAI, Anthropic, xAI, Moonshot AI and Alibaba's [Qwen3.8-Max](/models/qwen/qwen3.8-max). It beat its [Qwen3.6-27B](/models/qwen/qwen3.6-27b) predecessor and several competing models. WildClawBench recorded a 516-minute completion time but listed its cost as unavailable, leaving the economics of self-hosting outside the comparison.\n\nAlibaba's Qwen team has released Qwen3.8-27B as downloadable Apache 2.0 weights, giving developers a 27B multimodal model with native video input, adjustable reasoning and a context window that can be extended to 1 million tokens. [Qwen's announcement](https://qwen.ai/blog?id=qwen3.8&ref=runtimewire) and the [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-27B?ref=runtimewire) attribute the model to the Qwen team rather than an individual executive.\n\n### A deployment profile built around control\n\nQwen3.8-27B uses 64 layers and a 5,120-dimensional hidden representation. It is a native vision-language model that accepts text, images and video, including hour-scale video. The model has a native 262,144-token context window, which Alibaba says can be extended to 1 million tokens through context-scaling techniques.\n\nDevelopers can load the model through Hugging Face Transformers or serve it with vLLM, SGLang and TokenSpeed. Those standard serving paths make the released artifacts easier to test inside existing inference setups, although Alibaba does not disclose the model's training compute or provide measured hardware requirements for deployment.\n\nThinking is enabled by default. The model's `reasoning_effort`\n\nsetting supports `xhigh`\n\n, `medium`\n\nand `low`\n\nlevels, while `preserve_thinking`\n\nlets an application retain reasoning context between messages. Applications can disable thinking when per-turn latency matters more than extended analysis.\n\nAlibaba warns that lower reasoning effort can increase total time and token use on multi-step jobs when shallow analysis produces failures and retries. Agent operators therefore need to measure full task completion rather than relying on latency for a single response.\n\n### Alibaba's benchmark gains need outside replication\n\n[Alibaba's model card](https://huggingface.co/Qwen/Qwen3.8-27B?ref=runtimewire) reports Qwen3.8-27B at 61.7 on SWE-bench Pro, 73.0 on Terminal Bench 2.1 and 90.3 on LiveCodeBench v6. It also lists scores of 84.3 on OSWorld-Verified, 64.8 on WebArena-Verified and 81.9 on AndroidWorld.\n\nThose figures come from Alibaba's evaluations. Several tests used the Claude Code harness, a 256,000-token context window and baselines that Alibaba re-evaluated. QwenSWEBench and CoWorkBench are in-house tests. The results show gains over Qwen3.6-27B under Alibaba's listed setups, but they provide limited evidence about performance in another developer's agent stack.\n\nWildClawBench supplies a separate data point based on one harness. Its 48.0% score does not establish general model quality, and the missing cost figure prevents a direct price-performance comparison with hosted models. The result still narrows the claim developers can reasonably make: Qwen3.8-27B improved on its predecessor in this test while remaining below several larger or managed systems on end-to-end task completion.\n\nThat distinction matters for teams considering the downloadable weights. A 27B model can be useful when infrastructure control, data handling or runtime customization outweighs the highest available benchmark score. Each operator still has to test whether the model completes coding, browser and multimodal jobs reliably on the hardware and serving configuration it plans to use.\n\nAlibaba says an official hosted edition will be offered through [Qwen Cloud](https://www.qwencloud.com/models/qwen3.8-27b?ref=runtimewire) with a 1 million-token context window enabled by default and built-in tools. The repository does not provide an availability date or pricing. Until those details are published, the downloadable artifacts remain the version developers can pin and evaluate directly.", "url": "https://wpnews.pro/news/wildclawbench-ranks-alibaba-s-qwen3-8-27b-15th-on-agent-tasks", "canonical_source": "https://runtimewire.com/article/alibaba-releases-qwen3-8-27b-open-weights-agent-model", "published_at": "2026-08-17 04:50:18+00:00", "updated_at": "2026-08-17 05:10:45.935409+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "large-language-models"], "entities": ["Alibaba", "Qwen3.8-27B", "WildClawBench", "OpenClaw", "OpenAI", "Anthropic", "xAI", "Moonshot AI"], "alternates": {"html": "https://wpnews.pro/news/wildclawbench-ranks-alibaba-s-qwen3-8-27b-15th-on-agent-tasks", "markdown": "https://wpnews.pro/news/wildclawbench-ranks-alibaba-s-qwen3-8-27b-15th-on-agent-tasks.md", "text": "https://wpnews.pro/news/wildclawbench-ranks-alibaba-s-qwen3-8-27b-15th-on-agent-tasks.txt", "jsonld": "https://wpnews.pro/news/wildclawbench-ranks-alibaba-s-qwen3-8-27b-15th-on-agent-tasks.jsonld"}}