A paper proposes SkillGym, which turns human-written skills into 2,756 training environments and fine-tunes on 8,364 successful trajectories. It reports that Qwen3.5-35B-A3B in Claude Code then gains 19.10 points on Terminal-Bench 2.1 and reaches 51.47% on skill-assisted SkillsBench v1.1.
Qwen3.8-Omni-Flash is trained for long-horizon agent tasks across text, audio, and video, using the Qwen3.8-Next sparse mixture-of-experts design and a one-million-token context window. Two open-source frameworks, Qwen-MM-Plugins and Qwen-Live-Harness, add multimodal and real-time agent support.
OpenRouter says Kev is a LoRA adapter and pointer head on Qwen3.5-4B-Base, with weights in a Hugging Face repository, and thanked SiliconFlowAI for hosting.
The Jev-like classifier is finetuned on Qwen3.5 4B and priced at $0.042/M input and $0/M output. Together AI is also releasing a data recipe and a finetuning tutorial, and says training Tev1 cost $17.
OpenRouter now hosts Qwen3.8 Max Prime from Alibaba_Qwen, describing it as a higher-throughput variant of Qwen3.8 Max with the same 2.4T-parameter capabilities.
A Codex planner and creator turn each skill into tested terminal workflows, using a rubric taken from that skill's own quality criteria. Outcome-only RL moved Qwen3.8-27B from 49.4% to 54.1% on Terminal-Bench 2.1, while adding the rubric reached 50.1% but won more trajectory comparisons.
The discount applies to both Alibaba_Qwen models on Together AI serverless and runs through September 30. Together AI recommends Qwen3.7-Max for long-horizon agent work and Qwen3.8-Flash for high-volume, cost-sensitive workloads.
Victor M says the model turns long documents into small page images to save tokens, then retrieves the full text of only the pages relevant to a question.
Command Code has added 12 models, including Mimo V2.6 Pro Ultraspeed, GLM 5.3 FlashX, Qwen 3.8 Omni Flash, and Step 5 Preview. DeepSeek V4.1 Flash is included with extended usage, and Jev is free.
The lineup upgrades ASR, TTS, and Realtime and adds ASR-Next for audio understanding and TTS-Next for audio creation. Qwen puts the cuts at about 70% for TTS, about 85% for Realtime, and up to 95% for ASR.
ModelScope said Qwen has introduced RecreationBench, a 250-task benchmark for Hybrid Computer-Use Agents across Ubuntu, macOS, Windows, Android, and Web. Agents must explore a running app, recreate it in code, and pass programmatic tests plus VLM-based visual evaluation.
Alibaba Cloud highlighted a limited-time offer to try Qwen3.8-Flash for free on Token Harbor from Sep 20–27. Token Harbor said Qwen3.8 Flash is free for one week, with a 1M context window, and is live on its platform for those dates.
Alibaba Cloud said QwenWork can be given a one-time instruction to check multiple product pages on a set schedule, compare them with the previous week's snapshot, and send a report of price changes with timestamped screenshots.
Qwen highlighted a Hugging Apps live demo of Qwen-Image-2.1 on Hugging Face Spaces, saying one checkpoint covers generation and editing and can be tried in the browser with no setup. Hugging Apps described the model as a 7B-parameter native generation and editing model with up to 10 image references.
ComfyUI said Qwen-Image-2.1 is now supported, describing an open-weights 7B checkpoint for generation and editing. Qwen quoted that announcement and highlighted the ComfyUI support.
vLLM said Qwen-Image-2.1 from @Alibaba_Qwen has day-0 support in vLLM-Omni, describing a 7.1B DiT paired with Qwen3-VL-8B for generation, editing, and transparent images. Qwen thanked vLLM and said the one model is ready to serve.
Qwen announced Qwen-Image-2.1, saying the 7B model unifies image generation and editing and that its weights are now open. The company said it natively handles RGBA layers, supports up to 10 reference images, and offers precise local control.
