Alibaba's 2.4T Qwen3.8-Max is live in preview; weights still pending
A Kimi K3 rival you can call today: ~1M context, OpenAI/Anthropic-compatible, agent tools built in — but no open weights, no benchmarks, preview-only pricing.
Copy markdown
Callable today, weights aren't
The `qwen3.8-max-preview` endpoint is live now through Alibaba's Token Plan and Qwen Chat — a 2.4T-parameter sparse MoE with a ~984K-token context and up to 131K tokens of output. Open weights are still only “coming soon,” with no date or license, so you can query it but can't self-host it yet.
A drop-in for your OpenAI or Claude client
It exposes both OpenAI- and Anthropic-compatible APIs, so you can point an existing client at it by swapping the base URL and model id. Reasoning is tunable across low/high/xhigh, and built-in web search, code interpreter, web scraping and image search let you stand up an agent without wiring those tools yourself.
~10% pricing — but only during preview
There's no standalone per-token rate yet; access runs on subscription credits (individual plans roughly $6–$68/mo) that Alibaba says burn at about 10% of standard cost during preview. Cheap to prototype on now, but don't hard-code your economics around a promo that ends when the model stabilizes.
Benchmarks are vendor claims, not scores
Alibaba published no benchmark table, no methodology, and no active-parameter count. It reportedly claims the model is “second only to Claude Fable 5”; one independent software-architecture test put it at 80/100 versus Kimi K3's 83/100. Treat it as promising-but-unproven and run your own evals before shipping.
The open-model race gets a second front
This lands days after Moonshot's Kimi K3 (open weights due Jul 27) and reframes the frontier-open fight as a two-lab race out of China. If you're betting on downloadable weights, watch for Qwen3.8-Max's checkpoint drop — until then, K3 is still the one you can actually pull down.