Qwen3.8-Max and coding cowork
Qwen's newer releases are worth following because the ecosystem around Qwen models is becoming a default base for many open-model builders.
A model release used to mean a benchmark table. Qwen3.8-Max’s announcement contains almost none — instead it shows the model left alone for ten-plus days to build software, five days to reproduce and beat a research paper, and 24 hours to outscore 87% of human teams in a live contest. The unit of evaluation has quietly changed from “answers questions” to “finishes projects.”
Source
Qwen3.8-Max: A New Bar for Coding and Cowork
By: Qwen (Alibaba)
Published: Aug 02 2026
Type: Release announcement
I follow Qwen because its releases aren’t just about one model — the community fine-tunes, quantises, and deploys on top of them, so a strong base compounds through its derivatives. This post matters doubly: it’s the first time Alibaba is open-sourcing a Max-class model (weights promised a week after launch), built on the Qwen 3.5 architecture at 2.4 trillion parameters with 95B active.
Coding as multi-day autonomy, demonstrated three ways
The post’s framing: for a top model, coding now means “taking a real, multi-day project from an empty folder all the way to a finished result, on its own.” Three demonstrations, each with the same theme — the model self-evolves through feedback loops rather than following a fixed plan.
First, a 10-plus-day autonomous run building oh-my-cli, a self-evolving harness: requirements normalise into GitHub issues, an issue state machine dispatches agents (ready → leased → active), E2E tests and CI gate each merge, and the system feeds its own test failures back into new issues. After roughly 16 days it had accumulated 265 commits, 127 PRs, and 151 issues — and the full trace is public on GitHub, which is exactly the kind of verifiable artefact these announcements usually lack.
Second, research reproduction turned into research improvement. Given only the paper “Unified Data Selection for LLM Reasoning” and GPUs, the model spent ~37 hours rebuilding the pipeline from scratch and reproducing all six main findings (the paper’s selection method beats random data by +7.7% on AIME24). Then, over the next ~88 hours, it ran its own hypothesise-code-run-analyse loop through 18 improvement ideas across four rounds — and produced a method that beats the paper’s own approach by +2.7 points. About 7,600 lines of code, 33 rounds of GPU training, no human help.
Third, a live competition: the WWW2025 Multimodal Dialogue Intent Recognition Challenge on Alibaba’s Tianchi platform, 526 human teams, hard 24-hour limit. The model fine-tuned and ensembled text models (BERT, MacBERT, RoBERTa) with a vision-language model for screenshots, fused everything through calibrated weighted voting, and iterated 45 submissions from 0.60 accuracy to 0.853 — beating 458 of 526 teams.
Cowork: the messier half of the pitch
The “work” track argues the model handles multi-step, tool-heavy professional tasks. The technical core is scaling real-world RL across five harnesses (QwenWork, Claude Code, Codex, OpenClaw, Hermes) by scaling environments along three independent axes — task (single → multi-day), workspace (multi-file → heterogeneous), harness (category, version, skills) — plus a universal reward system that unifies execution checks, rubric-conditioned judging of text and rendered visual output, and an online balancer that keeps training batches stable. The showcases are aggressive: 1,284 relevant clauses surfaced from hundreds of documents in under an hour (vs a paralegal team’s week), an 8-screen banking prototype in one shot with zero revision rounds, a 26-dish menu from a hundred supply briefs with food-cost held at 33.8%, a browser-reconstructed seismic model of a 30-storey tower. Each is impressive; none is independently checkable.
The most concrete cowork capability is “Dynamic Workflows”: from six short factor descriptions, the model decomposed each into 50 research directions, dispatched ~330 sub-agents, ran ~6,000 backtests, and adapted the workflow mid-run — delivering quant factors with excess Sharpe ratios of 0.64–1.48 from a single conversation. A long-horizon section extends this to autonomous chip design (a GCD/RSA crypto accelerator verified bit-exact across bit-widths while minimising gate count). The through-line is orchestration as programming: planning logic frozen into reproducible programs that spawn and re-plan sub-agents.
My take
What the source reveals — intentionally or not — is that benchmark tables are being replaced by narratives of autonomy, and those are much harder to falsify. Qwen’s strongest evidence is precisely the part with a public artefact (the oh-my-cli repository); the profession showcases and quant results are stories with no external check. The right reading habit: weight claims by whether a trace exists. The second reveal is economic: open-sourcing a Max-class 2.4T model is an ecosystem play — every fine-tune and quantised derivative extends Qwen’s gravity long after the announcement cycle ends. The model matters less as a competitor to closed labs than as the base the open world will build on next.
Where this is useful
- Calibrating trust in “agentic” launch claims — look for public traces (repos, replays, submissions) and discount the rest.
- Choosing a base for fine-tuning once weights land: a Max-class open model changes the derivative ecosystem’s ceiling.
- Designing long-horizon agent evaluations: the issue-state-machine + CI-gated-merge loop is a reusable pattern for verifiable autonomy.
- Less useful for direct benchmark comparison — the post quotes almost no standard leaderboard numbers, which is itself the tell.
Apply this: when a lab claims autonomous multi-day capability, ask for the trace. If the artefact is public like oh-my-cli, study it; if it’s a story, file it under marketing until reproduced.
What I want to remember
Remember: Ecosystem gravity beats single benchmark wins — a model that thousands fine-tune and deploy compounds; one that only tops a leaderboard doesn’t. And autonomy claims deserve trust only when a public artefact exists to audit.