| Month | Releases | Share |
|---|---|---|
| 2025-09 | 1 | 1% |
| 2025-10 | 0 | 0% |
| 2025-11 | 0 | 0% |
| 2025-12 | 0 | 0% |
| 2026-01 | 0 | 0% |
| 2026-02 | 8 | 12% |
| 2026-03 | 2 | 3% |
| 2026-04 | 15 | 22% |
| 2026-05 | 1 | 1% |
| 2026-06 | 4 | 6% |
| 2026-07 | 0 | 0% |
| 2026-08 | 8 | 12% |
| 12-MONTH TOTAL | 39 | — |
Artificial Analysis added Qwen3.8 27B (medium) to its model catalog on 2026-08-21, measuring it at 44 on the AA Intelligence Index. The creator is listed as Alibaba. This is an independent evaluation, not a vendor announcement.
Artificial Analysis added Qwen3.8 27B (low) to its model catalog on 2026-08-21, measuring it at 43 on the AA Intelligence Index. The creator is listed as Alibaba. This is an independent evaluation, not a vendor announcement.
Qwen released Qwen3Guard, a safety classification model built on Qwen3 and fine-tuned for prompt and response moderation. It outputs risk levels and category classifications, and claims state-of-the-art results on major safety benchmarks per the vendor's tech report.
Qwen released Qwen-Image-Edit, a 20B image editing model built on Qwen-Image. It extends Qwen-Image's text rendering capability to editing tasks, and uses both Qwen2.5-VL (for visual semantic control) and the VAE Encoder (for visual appearance control) to support semantic and appearance editing.
Qwen released Qwen-Image, a 20B parameter MMDiT image foundation model with native text rendering capabilities. It handles complex text rendering including multi-line layouts, paragraph-level semantics, and fine-grained details, with support for alphabetic languages. Available via Qwen Chat for image generation.
Qwen published GSPO (Group Sequence Policy Optimization), a new reinforcement learning algorithm designed to address training instability in existing RL methods like GRPO. The paper identifies that existing algorithms exhibit severe instability during long training runs leading to irreversible model collapse, and proposes GSPO as a more stable alternative for scaling RL-based language model training.
Qwen released an updated version of Qwen-MT (qwen-mt-turbo) via its API, built on Qwen3 with reinforcement learning integration. The model supports 92 languages covering over 95% of the global population.
Qwen released Qwen3-Coder, a 480B-parameter Mixture-of-Experts model with 35B active parameters, achieving state-of-the-art results among open models on agentic coding, browser-use, and tool-use benchmarks, described as comparable to Claude Sonnet 4.
Qwen-TTS, available via the Qwen API, now supports Pekingese, Shanghainese, and Sichuanese dialects. The TTS model automatically adjusts prosody, pacing, and emotional inflection based on input text. Seven Chinese-English bilingual voices are available, including the three dialect-specific voices.
Qwen introduced VLo, a unified multimodal model that both understands images and generates high-quality visual content based on that understanding. It builds on the Qwen2.5 VL line, extending from perception-only to perception-plus-creation.