MarkTechPost chronicles Qwen's three and a half years: from 7B to a reported 2.4T open-weight model
According to a MarkTechPost chronicle dated October 4, 2026, Alibaba's Qwen went from an invitation-only trial in April 2023 to a reported 2.4 trillion parameter open-weight model, with licensing shifting from free commercial use under 100 million monthly active users to Apache 2.0 across multiple sizes, alongside MoE sparsity, hybrid linear attention, and parallel open and closed product lines.
What changed
According to the MarkTechPost article 'The Story of Qwen' published on October 4, 2026, author Asif Razzaq chronologically traces the three and a half year release trajectory of Alibaba's Qwen series: on April 7, 2023, Alibaba Cloud opened 'Tongyi Qianwen' trials to enterprise customers via invitation codes; on August 3 of the same year, Qwen-7B and Qwen-7B-Chat were open-sourced under a license permitting free commercial use below 100 million monthly active users; in 2024, Qwen1.5, Qwen2, and Qwen2.5 were released in rapid succession, with Qwen2 switching most sizes except 72B to Apache 2.0; in 2025, Qwen3 introduced 'hybrid thinking,' with the flagship Qwen3-235B-A22B fully under Apache 2.0, followed by Qwen3-Next-80B-A3B using a 3:1 hybrid of Gated DeltaNet linear attention and gated attention; in 2026, the report states Alibaba advanced multiple generations including Qwen3.5, Qwen3.6, and Qwen3.7 within half a year and released what it describes as a 2.4 trillion parameter open-weight model. The article also records team personnel changes and Alibaba's organizational restructuring of its AI business into a new group. The article's main text cuts off at the Qwen3.7 section, leaving details of some later nodes invisible, so this analysis relies only on the visible content.
What it means for developers
The immediate impact for developers appears first at the license boundary: the report shows early Qwen 7B used custom terms allowing free commercial use below 100 million monthly active users, while Qwen2 onward switched multiple sizes to Apache 2.0. Open weights and an open-source license are two different things: downloadable weights do not equal Apache 2.0, nor do they mean unconditional commercial use, redistribution, or use for training derivative models; the specific grant must be determined by the license file in the corresponding model repository, and this report's description cannot replace the original license text. Second is the split in deployment form: the report states that from Qwen3-Max-Preview onward some flagships are available only via API, and Qwen3.6-Plus is a closed hosted model, creating two engineering paths: 'self-hostable open weights' and 'hosted flagships accessible only via API.' Third are changes in architecture and cost assumptions: the report states Qwen3-Next-80B-A3B's training cost is about 10% of Qwen3-32B's and that inference throughput improves 10x beyond 32K tokens; such figures come from model cards and vendor statements, have not been reproduced by third parties, and can only serve as selection assumptions rather than commitments. Fourth, context specifications are inconsistent: the report mentions 32K for Qwen1.5 sizes, 128K for some Qwen2 instruct models, native 262K for the 2507 update of Qwen3-235B, and a claimed 1 million tokens for the hosted Qwen3.5-Plus; mixing these can lead to differences in actual truncation behavior and billing. Finally, the personnel and organizational changes in the report do not themselves constitute technical changes, but they suggest that release cadence and maintainership may shift, so long-cycle projects should budget for the compatibility cost of cross-version upgrades within the same series.
Integration and evaluation
The following is advisory analysis, not established fact. The first decision concerns licensing and distribution: if the report is accurate, different sizes within the same family may fall under custom terms and Apache 2.0 respectively, so before choosing self-hosting, check the license file for each target checkpoint individually rather than judging the series as a whole. The trigger is if you plan to embed the model in an external product, redistribute weights, or use it to train derivative models; for internal experimentation only, the license differences impose far fewer constraints. A verifiable criterion is: for each candidate checkpoint, confirm the license name, commercial use and redistribution terms, and record the date checked. The second decision concerns dense versus MoE trade-offs: the report states Qwen3.5-397B-A17B activates 17 billion of 397 billion total parameters, Qwen3.5-35B-A3B activates a smaller amount, and Qwen claims the latter exceeds the earlier Qwen3-235B-A22B-2507 on some measures. What is needed is selection by VRAM and throughput rather than total parameters: measure on target hardware the actual VRAM footprint determined by activated parameters, per-request latency, and throughput. The trigger is that when concurrency is high or VRAM is constrained, sparse models may be more cost-effective; if the task demands high stability in long-chain reasoning, the predictability of dense models may matter more. A verifiable criterion is comparing p95 latency and cost per thousand tokens on the same set of real requests, not just benchmark scores. The third decision concerns long context: since 32K, 128K, 262K, and 1M specifications coexist in the report, before enabling long context, confirm whether it is a native model capability or a hosted configuration. The trigger is if you need to process long documents or multi-turn agent trajectories; a verifiable criterion is testing retrieval accuracy and truncation points near the target length boundary.
Limits and open questions
The original source this article relies on cuts off at the Qwen3.7 section, and the names, licenses, and release dates of the 2.4 trillion parameter open-weight model and later nodes are incomplete in the visible text, so the above statements about that model are all in MarkTechPost's reporting terms and have not been independently verified. The versions mentioned in this report have no exact match in the current KimiSeek catalog; KimiSeek currently lists Qwen entries as Qwen3.8 Max, Qwen3.8 Flash, Qwen3.8 27B, Qwen3.8 2.4T A95B, Qwen3.7 Max, Qwen3.7 Plus, Qwen3.6 Plus, and two image models, which are not the same identifiers as the specific checkpoints of Qwen3.5, Qwen3.6, and Qwen3.7 described in the report; catalog entries should not be taken as evidence of availability for the reported versions. API availability mentioned in the report refers to supply on the vendor side. Performance and cost figures in the report, such as training cost ratios, throughput multiples, and benchmark comparisons, come from vendors or secondhand accounts and lack reproducible evaluation conditions, so they should not be the sole basis for procurement or migration decisions. License descriptions come from the chronicle's retelling and may differ from the formal terms in each model repository; the original license files should govern. Prices and billing change over time; refer to the corresponding platform's pricing page. Whether the personnel and organizational changes to the model affect long-term maintenance and release cadence cannot be determined from this source.