Qwen 3.8 Max

<p>Qwen 3.8 Max is Alibaba's frontier language model, announced on 3 August 2026 with the API live the same day. The architecture is a sparse mixture-of-experts: 2.4 trillion total parameters with roughly 95 billion active per token. Alibaba is positioning it around agentic work rather than chat — long-horizon tool use, multi-step coding, and native multimodal input handled by the same checkpoint rather than a separate vision variant. The headline benchmark in the announcement is TerminalBench at 86.6.</p><p>The number that actually matters here is on the pricing page, not the benchmark table: $2 per million input tokens and $6 per million output tokens. That is a deliberate undercut of closed frontier list pricing, and it lands in the part of the market where cost compounds hardest — agentic coding loops that burn tokens across dozens of tool calls per task. Pair that with the open-weights release Alibaba promised for the week after launch and the pitch is clear: frontier-tier agentic capability that you can buy cheaply now and, if the weights land as described, run yourself later.</p><p><strong>My read:</strong> treat the benchmark as a claim and the price as a fact. TerminalBench 86.6 is vendor-reported and has no independent replication yet; the open weights are a commitment with a date attached, not a download. Both are worth watching, but only one of them is verifiable today. The other caveat is scale — a 2.4T-parameter model, even at 95B active, is not something a small team self-hosts on a whim, so "open weights" here mostly benefits inference providers and well-resourced labs rather than individual developers. Note also that this is the model, not Alibaba's consumer <a href="/tools/qwen-app">Qwen App</a>; the two get conflated constantly in coverage.</p>