Alibaba Says Qwen3.8-Max-Preview Is Second Only to Fable 5; Benchmarks Haven't Confirmed It

Alibaba described Qwen3.8-Max-Preview as "one of the most powerful models available today, second only to Fable 5," while noting that the release is currently offered as a preview for developers. The company has released the model through its Token Plan, Qoder and QoderWork platforms, making it available for developers to try immediately. But it hasn't published a model card, licensing details or results from independent evaluators such as Artificial Analysis or LMArena, whose rankings are widely used to compare frontier AI models. That doesn't mean Alibaba's claim is wrong. It simply means it hasn't yet been independently verified. Bigger Doesn't Always Mean Better Alibaba says Qwen3.8 contains 2.4 trillion parameters, making it one of the largest AI models announced to date. AI researcher Simon Willison has repeatedly argued that parameter counts alone are no longer a reliable measure of model capability, particularly for mixture-of-experts (MoE) systems where only a fraction of the model is active during inference. Like most modern frontier models, Qwen3.8 is expected to use a mixture-of-experts (MoE) architecture, where only a small portion of the model is active for each request. Alibaba hasn't disclosed how many parameters are actually used during inference a figure that has a much greater impact on speed, cost and efficiency than the total parameter count. Without that information, it's difficult to compare Qwen3.8 directly with competitors based on size alone. What We Know and What We Don't | Confirmed | Still Unverified | |---|---| | API preview available now | "Second only to Fable 5" ranking | | 2.4T total parameters | Active parameter count | | Open-weight release promised | Release date and licence | | Compatible with OpenAI/Anthropic tool calling | Independent benchmark results | A Familiar Pattern The launch comes during one of the busiest periods in AI this year, following Grok 4.5, GPT-5.6 and Moonshot AI's Kimi K3 within just 11 days. Alibaba also owns roughly 36% of Moonshot AI, making the rapid succession of flagship launches notable even within its own ecosystem. Like Kimi K3, Qwen3.8 is being promoted as an eventual open-weight model. But for now, developers can access it only through Alibaba's hosted services. That's worth remembering because Alibaba followed a similar approach with Qwen 3.6 Max Preview, which remained API-only long after its initial announcement. What Developers Should Watch The good news is that Qwen3.8 isn't just a teaser. Developers can already test it through Alibaba's preview programme, evaluate its coding performance and verify compatibility with existing OpenAI- and Anthropic-style tool-calling workflows. What they can't yet verify are Alibaba's biggest performance claims. LMArena and Artificial Analysis have become widely used by developers because they evaluate models under standardized conditions, helping distinguish independently verified performance from vendor marketing claims. For now, Qwen3.8 looks like another serious contender in the frontier AI race. Whether it truly sits just behind Fable 5 or somewhere else on the leaderboard is a question that still needs evidence rather than marketing.