AI Daily · Sep 20, 2026: Qwen releases open-weight image model Qwen-Image-2.1 with editing and transparency
Alibaba Qwen released Qwen-Image-2.1, an open-weights 7B model unifying generation and editing with native RGBA transparency, 2K output, and multi-image editing; day-0 support in ComfyUI and vLLM-Omni. Meanwhile, StepFun unveiled its flagship agentic model Step 5 Preview focused on cost-efficient lo
Pick a date
-
Qwen releases open-weight image model Qwen-Image-2.1 with editing and transparency
Alibaba Qwen released Qwen-Image-2.1, an open-weights 7B model unifying generation and editing, with native RGBA transparency, 2K output, and multi-image editing from up to 10 references; day-0 support in ComfyUI and vLLM-Omni.
-
Jev decision model rivals LLMs in accuracy at lower cost and higher speed
OpenRouter tests show Jev, a lightweight decision model, matched popular LLMs on classification accuracy in a 200-case benchmark, ranked second-cheapest behind Qwen3.8 Flash, and was over 5x faster than next-fastest alternative.
-
FlappyBench test: Qwen and DeepSeek models score 9/10, Gemini 3.8 Flash lags
A developer-run FlappyBench test showed Qwen 3.8 Omni Flash and DeepSeek V4.1 Flash scoring 9/10 on a Flappy Bird build, while Gemini 3.8 Flash scored 5/10 despite multiple iterations and higher cost.
-
StepFun releases Step 5 Preview, a 600B-parameter agentic flagship model
StepFun released Step 5 Preview, a Mixture-of-Experts flagship model with 600B total and 27B active parameters, aimed at agentic tasks like software engineering, supporting up to 24-hour runs, now available on OpenRouter and HuggingFace.
-
Qoder halves Auto credit consumption rate from Sep 21
Starting Sep 21 00:00 UTC+8, each Auto call on Qoder consumes 0.5x the credits it did before, effectively halving credit usage.
-
Tencent and Jina AI release OCR models on Hugging Face
Tencent's WeVisDoc (2B, 4B) and Jina AI's jina-ocr-v1 are now on Hugging Face. On OmniDocBench v1.6, WeVisDoc-4B scores 95.38 overall, with NaviDC-OCR still leading.
-
OpenRouter launches Ori Eval for quantitative LLM comparison
Ori Eval is a quantitative tool to compare LLMs on your own workload, installed via one-line curl and run with `ori eval`.