Alibaba's Tongyi Qianwen team has updated its flagship model, Qwen3.8-Max, to the 0902 release, adding a fresh round of post-training focused on coding and professional office tasks. On Code Arena's frontend coding leaderboard, the new version scored 1691, taking first place and beating its predecessor by 22 points.
Alibaba also released a second figure that says more about the position it's aiming for: the new version's blended price averages about $5 per million tokens, while the models sitting second and third on the leaderboard price in at $20 and $12 respectively.
22 Points, $15 Apart
The gap at the top of the leaderboard is already thin. The 22-point gap between 1691 and the previous version is the kind of margin rival labs typically close within a week or two in human blind-scoring. What actually creates distance is the column on the right: run the same frontend task through the second-place model, and the bill quadruples.
According to public reporting, the new version's API pricing is $2 per million input tokens and $6 per million output tokens. Working backward from those numbers against the official blended rate of $5 (a rough estimate), output tokens would need to account for roughly three-quarters of total usage to land on that figure. That ratio isn't surprising for frontend coding tasks: the prompt is often just a few hundred lines of context, while the model spits out an entire page of component code — output naturally eats up the bulk of usage. The $5 blended rate reflects real "writing code" usage patterns; for input-heavy, output-light work like long-document summarization, the effective unit price would run even lower.
For teams in China, this pricing line gives a clear reference point. Over the past year, companies that have wired agentic coding into production have most often gotten stuck at the finance stage: whether the model writes correct code stopped being the question long ago; whether a bill for millions of tokens a day can get signed off is what decides things now. Matching the leaderboard leaders on score while cutting price to a quarter moves a procurement decision more than 22 extra points ever could.
2.4 Trillion Parameters, 95 Billion at Work
Behind the 0902 release sits a base model called Qwen3.8-2.4T-A95B. Total parameter count is 2.4 trillion, with roughly 95 billion active on any given forward pass — a classic sparse mixture-of-experts setup. Of 512 experts, 11 are activated per token: 10 routed experts plus one shared expert.
The architecture also leans toward long-running tasks. The model has 23 layers alternating Gated DeltaNet and Gated Attention, with a native context window of 262,144 tokens, extendable to roughly 1.01 million. Mixing linear attention with gated attention is a direction several domestic teams working on long-context models have converged on over the past six months, for a straightforward reason: bring the memory and inference overhead of long context down into a commercially viable range — otherwise a million-token context window is just an impressive number on a spec sheet.
The activation ratio also explains where that $5 figure comes from. The 2.4 trillion total parameters set the model's knowledge capacity; the 95 billion active parameters set how much "electricity" each call actually burns. Push sparsity this far, and per-inference cost naturally comes in well below a dense model of comparable scale.
Already Rolled Out Across Alibaba's Own Product Line
The new version is now callable directly through the Qwen AI platform's API service, and has also been wired into Qwen Office, Qoder, and the Qwen app. Enterprise customers, developers, and everyday users all got access in one go, with no staged rollout window.
Alibaba is positioning this release as "better suited to enterprises' real, complex tasks, scientific research, and long-horizon work," saying the model has "reset the ceiling for global models" in multi-step reasoning, tool calling, and end-to-end app generation. The second half of that is the vendor's own framing, but the "long-horizon work" claim in the first half does line up with the two technical choices behind it — a 1.01-million-token context and sparse activation. Being able to hold an entire codebase in memory at once without blowing through the cost budget is what it actually takes for a model to keep working for hours at a stretch.
The naming convention also signals the pace Alibaba is setting. Using a date as the sub-version number for Qwen3.8-Max shows that on its closed-source Max line, Alibaba is running small, fast iterations — the model generation stays fixed while capability keeps stacking on top. Ever since 3.6-Max went closed-source, the Max series has split roles with the open-source Flash series: one chases leaderboard rankings and enterprise budgets, the other spreads the developer ecosystem. By pushing price down to a quarter of the leaderboard's top bids, this 0902 release tightens that division of labor by one more notch.
Sources: Alibaba Tongyi Qianwen official announcement, CocoLoop, Code Arena frontend coding leaderboard, ITHome, Hugging Face model card; the 1691 score and 22-point gain, the $5/$20/$12 blended per-million-token pricing, and the 2.4-trillion total versus 95-billion active parameter figures were each checked individually.