Alibaba’s Qwen team announced QwenGyre on 12 October 2026, a framework for training AI agents with reinforcement learning on tasks that run for hours, according to its post on X and the paper the post links. The paper, submitted to arXiv on 27 September 2026 by researchers at Alibaba Token Hub, says a single such task can span hundreds of model and tool interactions and nearly 1M tokens.
Applied to Alibaba’s flagship Qwen 3.8 2.4T model with about 700K tokens per rollout, the paper says QwenGyre lifted the model’s NL2RepoBench pass rate from 52.5% to 58.5% in 48 training steps, and finished those steps in 75.42 hours, against 134.55 hours for an Async baseline and 91.47 hours for a Colocate baseline.
Business Pill · THE MODEL AND THE HARNESS
A one-minute explainer of the difference between an AI model and the harness that runs it. It teaches the general idea only and says nothing about any organisation in this story.
The key insight: As we read it, QwenGyre is about the cost of training agents rather than running them: the paper says GPUs sit idle while hours-long tasks finish, and the speedups it reports come from putting that idle capacity to work on training without stopping the tasks still running.
The paper describes two problems. In its words, at these lengths “severe execution variance and prolonged rollout delays cause massive GPU idling”, and branching runs create so many overlapping training samples that they slow training down.
It says the common Colocate design switches one shared GPU pool between running the agent and training on the results, so training waits for the slowest task to finish. The Async design splits GPUs into fixed pools, which the paper says leaves training GPUs waiting for data while neither side can borrow the other’s spare capacity.
The paper gives the scale of the problem. On its 122B-parameter Qwen 3.6 tests, a rollout execution took 1.93 hours on average and a query 2.96 hours. On Qwen 3.8 2.4T, it says 61.1% of queries with complete duration records took at least four hours.
Hours to complete 48 reinforcement learning training steps on Qwen 3.8 2.4T on NL2RepoBench, as reported in Alibaba’s QwenGyre paper: 75.42 hours with QwenGyre, 91.47 hours with a Colocate baseline and 134.55 hours with an Async baseline. Alibaba’s own experiment.
It launches with two core designs: ๐ฅณ – Elastic Scheduler: reallocates GPUs between rollout & training as demand shiftsโlive executions never interrupted. Up to 1.85ร speedup overโฆ pic.twitter.com/wMDoJOuP2j
The Qwen team’s announcement on X, 12 October 2026. It introduces QwenGyre, names its two parts (the Elastic Scheduler and the Trajectory Processor), gives the speedups over the Colocate and Async baselines and the NL2RepoBench result on Qwen3.8-2.4T-A95B, and links the paper. This piece takes its figures from that post and the paper.
What QwenGyre Changes
The Qwen post names two parts. An Elastic Scheduler moves GPUs between rollout and training as demand shifts, with “live executions never interrupted”, and a Trajectory Processor turns branching runs, such as context compaction, sub-agents and retries, into shared trajectory trees. “Partial progress after timeout still gets scored,” the post says.
The paper says GPUs are grouped into cells that can join training one at a time while other cells keep serving the running tasks, and reports average switching times of 8.52 seconds from rollout to training and 3.46 seconds back.
On the data side, the paper says QwenGyre keeps at most five trajectories per execution by default. In its ablation, a cap of five matched the training scores of using every trajectory while cutting forward and backward time per step by 25.2%.
Trained Through Claude Code
The paper says all three training workloads used Claude Code 2.1.220, Anthropic’s coding agent, to run shell commands and edit files in containers. It treats the agent harness as a black box: the framework records model calls and task results without changing how the harness runs.
Its reason, as the paper puts it, is that the most capable harnesses, “such as Claude Code and Codex, are closed and evolve rapidly”, and that rebuilding their logic inside a training framework would train the model under a different harness from the one it is deployed with.
The Numbers, All From Alibaba’s Paper
On Qwen 3.8 2.4T, the paper reports end-to-end speedups of 1.78x over Async and 1.21x over Colocate across 48 training steps, using four cells of 48 nodes each. The Qwen post puts the time saved at about 44% against Async, which is consistent with the paper’s 75.42 and 134.55 hours.
On the smaller Qwen 3.6 122B model, the paper reports 1.42x to 1.53x speedups over Async and 1.36x to 1.47x over Colocate on NL2RepoBench, and 24-step speedups of 1.38x to 1.57x over Async and 1.61x to 1.85x over Colocate on DeepSWE and TerminalBench.
The paper says these comparisons used equal GPU budgets and matched staleness settings, and that training scores were comparable across methods. The 1.85x figure in Alibaba’s post is the best Colocate result in the paper, and 1.78x the best Async result.
The Structural Read
As we read it, the harness choice is the detail to watch: the paper trains Qwen through Claude Code and says rebuilding closed harnesses inside a training framework would train the model under a different harness from the one it is deployed with.
As we read it, the baseline matters for the headline numbers: on Qwen 3.8 2.4T the paper reports 1.78x against Async but 1.21x against Colocate.
As we read it, partial credit is part of the design: the post says progress after a timeout still gets scored, and the paper reports that 61.1% of Qwen 3.8 2.4T queries with complete records ran for at least four hours.
Qwen team (@Alibaba_Qwen), on X, 12 October 2026
“Partial progress after timeout still gets scored.”
Three Implications
LABS TRAINING AGENTS The paper says QwenGyre moves GPUs between running agents and training them as demand shifts, without interrupting live executions.
TEAMS BUILDING ON HARNESSES The paper describes training Qwen through an unmodified Claude Code harness, recording only model calls and task results.
READERS COMPARING CLAIMS Every speedup and score here comes from Alibaba’s own experiments, and the paper says much smaller GPU budgets remain unevaluated.
The framework puts it this way: “The agentic surface — the harness — is becoming where value pools in the entire AI stack.”
As we read it, QwenGyre points the same way from the training side: Alibaba’s paper treats the harness as fixed and builds its training system around it, running Claude Code unmodified while the framework records model calls and task results.
Figures from Alibaba’s QwenGyre paper and the Qwen post of 12 October 2026: Qwen 3.8 2.4T’s NL2RepoBench pass rate rose from 52.5% to 58.5% over 48 training steps with about 700K tokens per rollout; end-to-end speedups of 1.78x over Async and 1.21x over Colocate, on four cells of 48 nodes each. Alibaba’s own experiments.
What Is Not Established
Every result here comes from Alibaba’s own experiments; we have not seen independent replication. The paper’s main text, which we read, makes no statement about releasing QwenGyre’s code, and the post links only the paper. The paper itself says efficiency under much smaller GPU budgets remains unevaluated, and that the effect of its data ordering on learning quality has not been isolated. QwenGyre is a training system, not a product with a price.
Business Engineer Framework
The Agentic Harness War
A Business Engineer framework on why the harness, not the model, is where agentic value is pooling.
Alibaba’s Qwen team has published QwenGyre, a training framework it says finished 48 reinforcement learning steps on Qwen 3.8 2.4T in 75.42 hours instead of 134.55 with an Async setup, while raising the model’s NL2RepoBench score from 52.5% to 58.5%, on its own tests.
A note on sourcing. On 12 October 2026 we read the Qwen team’s post on X announcing QwenGyre, through api.fxtwitter.com, and the paper it links (arXiv 2609.33848, submitted 27 September 2026), including its results tables and limitations section. Every result is from Alibaba’s own experiments. We did not read any independent evaluation or code repository. Nothing here is investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
