In the current phase of large model development, progress has increasingly centered on systems that can sustain coherent activity across extended time frames rather than isolated responses.
Researchers and developers have spent the past year refining architectures and training methods that support multi-day coding sessions, multi-step professional workflows, and closed-loop optimization over hundreds of interactions.
Against that backdrop, the Qwen team at Alibaba has published details on its latest flagship system.
'Qwen3.8-Max' is described in a blog post as a 2.4-trillion-parameter model with 95 billion active parameters.
It is positioned for coding tasks that run for many days without continuous human oversight, for production-style work across a range of professions, and for long-horizon problems that require ongoing planning and correction.
The model is already available through QwenCloud and the associated API. Pricing is listed at $2.00 per million input tokens, $6.00 per million output tokens, and $0.25 per million tokens for implicit caching. Open weights for Qwen3.8-Max are scheduled for release in the coming week, together with a smaller 27-billion-parameter variant.
One documented coding episode involved the construction of a command-line project called oh-my-cli.
Over roughly 16 days, the model generated 265 commits, 127 pull requests, and 151 issues. It maintained an internal issue state machine, dispatched work, monitored progress, and triggered build, unit, end-to-end, and desktop lifecycle tests after each change.
The resulting repository is public.
In a separate exercise the model reproduced and then modified a research paper on data selection for language-model reasoning. It wrote approximately 7,600 lines of code, completed 33 GPU training runs, recovered the paper’s main findings, and later improved the AIME24 score by an additional 2.7 points through iterative idea generation.
Across professional domains the same system produced concrete outputs under realistic constraints. It reviewed hundreds of compliance clauses in under an hour, generated an eight-screen interface prototype in a single pass, assembled a twenty-six-item restaurant menu with caloric and cost data, reconstructed a thirty-story seismic model inside a browser, converted a two-dimensional rehabilitation form into an interactive three-dimensional demonstration, and summarized 8,400 sports possessions into a tactical report.
In quantitative finance it designed an ETF-rotation strategy from a one-line prompt, mined factors with hundreds of parallel sub-agents, and returned excess Sharpe ratios between 0.64 and 1.48.
Longer experiments tested sustained planning. In a chip-design task the model optimized a GCD/RSA cryptographic accelerator through more than 500 turns of RTL, simulation, synthesis, and layout. Gate count fell from 8,298 to 678 and die area from 106 by 106 micrometers to 46 by 46 micrometers while meeting a 500 MHz timing target. In a 365-day e-commerce simulation that began with ¥100,000 of capital across multiple store types and suppliers, the model finished with a balance of 416,252 yuan after managing product selection, pricing, inventory, negotiation, and fraud detection.
The architecture incorporates vision as a continuous feedback signal rather than a one-time input.
Screenshots, documents, and video frames can be inspected during execution so that layout errors, object mismatches, or process deviations trigger automatic correction.
Community discussion on the day of the announcement focused largely on the forthcoming open-weight releases, particularly the 27-billion-parameter model that many local-deployment users have requested.
Independent observers have begun comparing the listed API rates with those of contemporaneous systems and examining early benchmark numbers on coding and tool-use suites, though comprehensive third-party evaluations are still emerging.
The combination of scale, multi-day autonomy, and scheduled weight release places the model within the ongoing shift toward systems that treat extended, self-correcting work as a primary capability rather than an afterthought.
The announcement reflects a noticeable shift in how leading AI labs are defining progress.
Rather than focusing primarily on benchmark scores or single-turn reasoning, the emphasis is increasingly on persistence: systems that can plan, execute, monitor their own work, recover from mistakes, and continue making progress over days or even weeks.
That evolution mirrors how human professionals approach complex projects, where success depends less on producing a single correct answer than on sustaining productive work over long periods.
As more open-weight frontier models become available, competition is likely to move beyond raw intelligence toward reliability, efficiency, and autonomous execution. The next wave of comparisons may no longer ask which model writes the best code in one prompt, but which one can be trusted to manage an entire project from start to finish with minimal human intervention.
