Summary
- Alibaba's Qwen team released Qwen3.8-Max-0902 on September 2. It keeps the 2.4-trillion-parameter architecture and 1-million-token context window, retrained specifically for coding and collaboration tasks.
- Its TerminalBench 3.0 score jumped from 11.3 to 29.0, but still trails Claude Opus 5's 42.7.
Qwen3.8-Max Upgraded to the 0902 Version
Alibaba's Qwen team announced Qwen3.8-Max-0902, a new version of Qwen3.8-Max, via its official X account on September 2. The company kept the same 2.4-trillion-parameter mixture-of-experts (MoE) architecture and 1-million-token context window, but added further training focused on coding and collaboration (Coding & Cowork) tasks, saying it improves performance on complex enterprise workflows and scientific research problems.
To put that in context: Qwen3.8-Max is Alibaba's flagship language model, first unveiled earlier last month, and it briefly topped the Agentic Index in early August. This 0902 release keeps that base model intact and adds a follow-up training pass tuned specifically for coding and office collaboration tasks.
Related article: Qwen3.8-Max tops the Agentic Index, ranks 5th in overall intelligence index
Coding Benchmarks: Big Gains, Gap Remains
The comparison table Alibaba released shows most coding-related metrics climbing sharply over the previous version. Compared with Claude Opus 5, though, it still falls short on several fronts.
| Benchmark | 0902 | Previous version | Claude Opus 5 | Fable5 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal agent coding (TerminalBench 3.0) | 29.0 | 11.3 | 42.7 | 34.0 | 34.6 |
| Repository-level code generation (NL2Repo-Bench) | 64.9 | 55.9 | 72.3 | 70.2 | -- |
| Complex real-world software engineering (QwenSWEBench V2) | 70.0 | 55.1 | 68.0 | 67.1 | -- |
TerminalBench 3.0 in particular nearly tripled, from 11.3 to 29.0, but it's still more than 13 points behind Claude Opus 5's 42.7. On the other hand, on Alibaba's own QwenSWEBench V2 benchmark, it actually surpassed Claude Opus 5.
Office Task Agents: Ahead of Claude on WorkArena
Metrics measuring long-horizon office work and job-performance capability paint a more balanced picture.
| Benchmark | 0902 | Previous version | Claude Opus 5 | Fable5 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Long-horizon office work (CoWorkBench) | 76.1 | 74.8 | 79.6 | 75.9 | 71.5 |
| Professional job tasks (JobBench) | 64.0 | 53.4 | 67.8 | 57.4 | 45.4 |
| Expert-graded collaboration index (WorkArena Elo) | 1468 | 1348 | 1437 | -- | 1482 |
WorkArena is an Elo score derived from human grading across 145 tasks in 29 industries. The 0902 version scored 1468 on this metric, edging out Claude Opus 5's 1437, though it still falls short of GPT-5.6 Sol's 1482. The grading methodology is detailed in a separate footnote, which notes that the automation-workflow metric, Automation Bench v1.0.6, was measured after incorporating Automation Bench PR #16.
Multimodal and Embodied Intelligence: Strong on Robotic Perception Benchmarks
On multimodal metrics that test understanding of images and video together, the 0902 version opened a noticeable lead.
| Benchmark | 0902 | Previous version | Claude Opus 5 | Fable5 | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Embodied intelligence (ERQA) | 78.3 | 77.8 | 67.3 | 70.0 | 70.0 |
| General visual reasoning (BabyVision) | 93.8 | 91.3 | -- | 90.5 | 88.9 |
| STEM visual reasoning (MMMU-Pro) | 82.7 | 82.3 | 82.6 | 81.2 | 83.0 |
Editor's Take
The "0902" version name itself reveals Alibaba's release strategy. Rather than shipping an entirely new model, the company kept the same parameter count and context architecture and simply layered on follow-up training for specific task categories — a move that saves the cost of retraining from scratch while targeting exactly the areas where rival models have an edge. Compared with the original launch last month, which didn't even disclose the active parameter count, this release at least shows clear numbers for what improved and where.
Applying this kind of retraining-based update to real work tends to lead to the same conclusion every time: when a model's score jumps sharply on one benchmark, the improvement is only reliably felt in the task category that benchmark represents, while everything else often feels no different from the previous version. Here too, Claude Opus 5 remains the safer bet for long-horizon autonomous coding work in a terminal, but for internal document work or repetitive office automation, the WorkArena numbers suggest Qwen could be a solid, cost-effective alternative.
The takeaway for companies is clear enough: when deploying a coding agent, don't try to cover every task with a single model. It makes more sense at this stage to split assignments by category — keeping a different model for long-horizon terminal work while testing Qwen3.8-Max-0902 for office automation and document processing. In the coming weeks, Anthropic or OpenAI could well roll out similarly specialized retrained versions of their own — a sign that frontier-model competition is shifting from full retraining runs to pinpoint updates targeted at specific task categories.
Comments