Chinese technology giant Alibaba has unveiled Qwen3.8-Max, which the company describes as its largest and most capable AI model to date, arriving close on the heels of a similarly large model from domestic rival Moonshot AI. Qwen3.8-Max carries 2.4 trillion parameters, the numerical settings a model learns from data and uses to recognise patterns, generate answers, and carry out tasks, compared to the 2.8 trillion parameters in Moonshot's Kimi K3, which launched shortly before it. According to Bloomberg, Qwen3.8-Max ranks higher than Kimi K3 on some benchmarks, and Alibaba has said it plans to release the model's weights for public download, positioning the model as 'second only' to Anthropic's Claude Fable 5 in its own internal evaluations.
What Alibaba's benchmark charts show
Alibaba has published a series of self-reported benchmark comparisons pitting Qwen3.8-Max against its own predecessors, Qwen 3.7 Max and Qwen 3.7 Plus, as well as against Anthropic's Claude Opus 4.8 and Fable 5, Google's Gemini 3.1 Pro, and OpenAI's GPT-5.6 Sol. On software engineering and agentic coding benchmarks, Qwen3.8-Max scored 67.7 on SWE-Pro and 86.6 on TerminalBench-2.1, the latter ahead of every other model in the comparison including Fable 5's 84.6. On PaperBench, a research reproduction benchmark, Qwen3.8-Max posted a leading score of 93.0, ahead of Fable 5 at 88.8. However, on FrontierSWE, a separate software engineering benchmark, Qwen3.8-Max trailed both Opus 4.8 and Fable 5, scoring 73.5 against 88.8 for Fable 5.
On tasks measuring professional workflow and agentic capability, Qwen3.8-Max scored 74.8 on CoWorkBench and 53.4 on JobBench, both ahead of its predecessor Qwen 3.7 Max but behind Fable 5's scores of 75.9 and 57.4 respectively on the same benchmarks. On Agents' Last Exam, a benchmark for long-horizon, high-value tasks, Qwen3.8-Max scored 52.4, narrowly behind GPT-5.6 Sol's 53.6.
On visual and multimodal reasoning benchmarks, Qwen3.8-Max posted strong results on BabyVision and CharXiv, chart-reasoning tasks, though it trailed Fable 5 on both when Python-assisted scores were compared. On embodied reasoning benchmark ERQA, Qwen3.8-Max led the field with a score of 77.8, ahead of Fable 5's 70.0. On agentic computer and mobile use benchmarks, OSWorld-Verified and MobileWorld, Qwen3.8-Max scored 86.1 and 77.8 respectively, again leading Fable 5's scores of 85.0 and 85.5 on the former but trailing on the latter.
Alibaba's broader AI strategy
Alibaba has said the new model uses a mixture-of-experts design, an architecture that divides tasks among specialised internal components rather than relying on a single unified system. According to Bloomberg, the release marks the latest move in an intensifying race among Chinese AI developers that also includes Moonshot, DeepSeek, and ByteDance, with both Qwen3.8-Max and Kimi K3 being positioned as open-weight releases that developers can freely download and build upon, in contrast to the closed access models offered by leading US labs such as Anthropic and OpenAI.
