Video: "New Qwen3.8 Max is INSANE!" by Julian Goldie on YouTube.
What was announced at WAIC
The World Artificial Intelligence Conference ran in Shanghai in late July 2026 and Alibaba used it to announce Qwen 3.8 Max, the largest model in the Qwen 3 family by parameter count. The 2.4 trillion figure refers to total parameters; in a mixture-of-experts architecture, only a fraction of those are active on any given inference pass — which is why such large models can run at useful speeds rather than requiring proportionally more compute for every call.
Qwen 3.8 Max follows the Qwen 3 series that began with smaller and mid-sized models released earlier in 2026. The naming suggests it sits at the top of that family, though Alibaba has not publicly detailed the exact active-parameter count or which tasks trigger which expert subsets.
The benchmark claims
Alibaba placed Qwen 3.8 Max second overall on the internal evaluations they ran — behind Fable 5 and ahead of other frontier models. This is a significant claim and the appropriate response to it is scepticism until independent benchmarks appear.
Vendor benchmarks are common at announcement events and frequently measure tasks the announcing company's model was optimised for. Independent evaluations on Chatbot Arena, HELM, and similar platforms give a more reliable picture. As of the time of this article, independent Qwen 3.8 Max results had not yet appeared in public leaderboards. Julian noted the claim in his video coverage; he did not say it had been verified externally.
That said, the Qwen family has a credible track record. Earlier Qwen 3 releases performed above expectations on independent tests. It would not be surprising if the Max variant turns out to be genuinely competitive at the frontier.
What the model handles
Qwen 3.8 Max is multimodal in the broader sense. A single call can take text, images, video, and document inputs, rather than requiring separate models for each modality. This matters for agent use cases: an agent processing a PDF report, pulling a screenshot, and then writing a summary can do all of that within one model context rather than routing between specialised tools.
Alibaba's own platforms — Qwen Chat and the KODA productivity suite, including KOD Work — were confirmed to be receiving Qwen 3.8 Max in the same announcement. Enterprise access through API is expected, though pricing details had not been published at announcement time.
What this means for open-source AI models
Qwen models have been open-source or open-weight, and that matters here. If Qwen 3.8 Max follows the pattern of its predecessors, the weights will eventually be released under a permissive licence that allows commercial use and local deployment. That would put a model claiming frontier performance in the hands of anyone with the hardware to run it — or in Hermes Agent's model list, where Kimi K3 and earlier Qwen variants already appear.
The broader direction is worth noting. A year ago, the gap between the best closed US models and the best open-source alternatives was wide enough that using open models for agent work meant accepting meaningful capability trade-offs. That gap has narrowed sharply. Chinese labs — Alibaba with Qwen, DeepSeek, and others — have contributed significantly to that narrowing. Qwen 3.8 Max, if the benchmark claims hold up, is another step in the same direction.
What to wait for before acting on this
The announcement is real; the claims need independent confirmation. Before switching an AI agent's model config to Qwen 3.8 Max, it is worth waiting for: independent benchmark results on reasoning, coding, and instruction-following tasks; API pricing and availability; and clarity on whether weights will be publicly released and under what terms. None of those were confirmed at WAIC.
If the independent numbers are good, this matters for anyone running Hermes Agent or similar stacks — a capable open model available via OpenRouter or as local weights would lower the cost of sustained agent work significantly.
Where this connects to NordSys
The AI agents we configure for clients can be updated to use the best-available model as the landscape changes. If Qwen 3.8 Max proves out, we will add it to the options we offer. If you want an agent running your business tasks without having to track every model release yourself, our AI Agents service handles the configuration and keeps it current. From £6 a day.
See our AI Agents →