Alibaba debuts Qwen3.8-Max model with 2.4T parameters
SiliconANGLE Maria Deutscher ● Covered by 10 sources
Alibaba just dropped Qwen3.8-Max, a massive 2.4 trillion parameter AI model. It handles million-token prompts and nearly matches Anthropic's best on coding tests.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Alibaba isn't slowing down its LLM arms race. The company's latest release, Qwen3.8-Max, packs 2.4 trillion parameters — roughly seven times the size of Qwen3.5, which only came out in February. Despite that bulk, the model activates a comparatively modest 95 billion parameters per query, a sign that Alibaba is still leaning on sparse, efficiency-minded architecture rather than brute force alone.
What really stands out is the context window. Qwen3.8-Max can chew through prompts up to one million tokens, which Alibaba says translates to more than 200 pages of text or roughly 100 hours of video in a single request. Responses can run up to 131,000 tokens. That kind of scale hints strongly at a Gated DeltaNet-style attention mechanism, the same linear-scaling approach used in Qwen3.5 and Qwen3.6. Alibaba hasn't confirmed the architecture for this release, but the math only works if memory and compute demands grow linearly with prompt size rather than quadratically, which is exactly what Nvidia's gated update and delta rule techniques were designed to fix.
On benchmarks, the numbers are hard to ignore. Qwen3.8-Max reportedly finished a 16-day autonomous coding project without human intervention and pushed through a chip-design task spanning more than 500 steps. On the Frontend Code Arena benchmark, it scored 1,668, landing just 37 points shy of Anthropic's top Claude Opus 5 configuration and ahead of over a dozen rivals, including Meta's Muse Spark 1.1.
Alibaba is rolling the model out on its own cloud platform first, with a plan to open-source both Qwen3.8-Max and a leaner Qwen3.8-27B variant next week. That timing matters. Open-sourcing a model this close to frontier-lab performance puts real pressure on Western labs that have leaned on closed weights as a competitive moat.
My take — AI-written commentary, not fact-checked reporting
Another quarter, another Chinese lab quietly closing the gap with the best closed-source American models, and this time they're planning to just give it away. Anyone still betting that closed weights are a durable moat should be paying attention to how fast that argument keeps aging. If Alibaba's benchmark claims hold up under independent testing, this is less about one model and more about the shrinking shelf life of frontier exclusivity.
Read more about this at: SiliconANGLE