Qwen3.8 Preview: 2.4 T Parameters and Upcoming Open Weights

Qwen3.8’s first official details are brief but highlight two key points: a total parameter count of 2.4 T and the promise of “open weights soon.” The preview model, identified as qwen3.8-max-preview, is already available for users to test before the final release.

Confirmed Information

Model name: qwen3.8-max-preview Parameter size: 2.4 T

Capability tags: inference, visual understanding, text generation

Status: preview, with ongoing updates

Access: Token Plan, Qoder, QoderWork

Weights: officially “open soon,” but no download link or license disclosed yet

The preview is listed under inference, visual understanding, and text generation. In tools compatible with the Anthropic protocol, the reasoning mode can be set to reasoning_effort with options xhigh, high, and low. It is already integrated into developer and agent toolchains, not limited to a simple chat demo.

How to Try It in China – Pricing Details

Chinese users can access the model through the Qwen AI platform’s Token Plan, which uses credits to cover model usage, web search, web‑scraping, and code interpreter tools. The current limited‑time pricing for personal plans is:

Lite: 39 CNY/month for 700 credits (5‑hour quota) and 2,500 credits (7‑day quota)

Standard: 139 CNY/month for 3,000 credits (5‑hour) and 10,000 credits (7‑day)

Pro: 499 CNY/month for 12,000 credits (5‑hour) and 40,000 credits (7‑day)

During the promotional period, credit consumption is charged at a 10 % rate, with an additional 20 % discount between 22:00 and 08:00 the next day, effectively making it 0.2 × the standard price. Credits differ from tokens, so direct cost comparison with typical API pricing is not straightforward.

Claims vs. Evidence

The announcement boldly states that Qwen3.8’s overall capability is “only behind Fable 5,” positioning it alongside leading frontier models. However, the official release provides no benchmark, test set, comparative model version, or model card, making the claim unverifiable. Critics have asked for a smaller, locally‑runnable version and have noted the lack of concrete performance data.

Deployment Realities

While the total parameter count is 2.4 T, hardware requirements are unspecified. Details such as activation‑per‑inference parameter count, quantization strategy, context length, and GPU memory needs are absent. Earlier Qwen series used MoE architectures, but without a model card it is unsafe to assume the same design for Qwen3.8.

Current best practice is to experiment with the low‑cost preview, evaluating stability for coding, long‑running tasks, and visual inputs. Production migration should wait until benchmarks, a full model card, licensing terms, and weight download links are released.

Key Takeaways

Qwen3.8 offers a high‑parameter preview with flexible reasoning modes and an attractive Chinese pricing scheme, but the lack of performance evidence and open‑weight details means it remains a preview rather than a production‑ready solution.