Skip to content
Daily BriefNewsDaily Brief

4 AI stories from September 21, 2026: Alibaba Qwen-Image-2.1 drops Apache license, Qwen3.8-Omni-Flash cuts audio costs 98 percent, StepFun Step 5 Preview API opens at $1 per million tokens, and California orders AI kill switch review

· by Pondero Newsdesk · 4 stories

AI news daily brief: 2026-09-21

Four stories today: two model releases from Alibaba (Qwen-Image-2.1 and Qwen3.8-Omni-Flash), one large-scale agent model from StepFun, and one regulatory development out of California.

Alibaba Ships Qwen-Image-2.1, a 7B Image Model That Drops Apache 2.0 for a Research-Only License

Alibaba's Qwen team published Qwen-Image-2.1 weights on HuggingFace and ModelScope on September 20, 2026 under a research-only license that replaces the Apache 2.0 terms of earlier Qwen-Image releases, per the model card. The architecture pairs a 7B single-stream diffusion transformer image backbone with a Qwen3-VL 8B text encoder and a 64-channel RGBA VAE, producing native 2048x2048 outputs in 40 sampling steps. It accepts up to 10 reference images for guided generation and supports mask- and circle-based local edits through a prompt-rewriter checkpoint. Per Alibaba, the model targets quality competitive with Google's Nano Banana Pro at lower compute cost. The Qwen Research License Agreement is dated September 20; commercial deployment now requires a separate agreement with Alibaba. A thread on the model page has already formed requesting a return to open terms. Whether community pressure reverses the decision is the variable to watch.

Full story: Qwen-Image-2.1 and its research-only license

Alibaba Releases Qwen3.8-Omni-Flash with 98 Percent Lower Audio Cost and a 1 Million-Token Context Window

Qwen3.8-Omni-Flash became available through Alibaba's API on September 18, 2026 at $0.15 per million input tokens with a 1 million-token context ceiling, per TechNode. The model accepts text, images, audio, and video through the Chat Completions and Responses APIs in a single call and returns text output. Output costs $0.47 per million tokens. Per Alibaba, an hour of audio input costs 98 percent less than on the predecessor Qwen3.5-Omni-Plus; combined audio and video, 93 percent less. Per Alibaba, average scores across 30 evaluations improved by more than 26 percent compared with Qwen3.5-Omni-Plus. Deployments cover Beijing, Singapore, Hong Kong, Tokyo, Frankfurt, and Virginia. For operators running speech-to-action pipelines on higher-cost multimodal endpoints, there is a direct pricing alternative to evaluate now. Whether third-party benchmarks confirm the performance claims against competing endpoints is the open question.

Full story: Qwen3.8-Omni-Flash and the 98-percent audio cost reduction

StepFun Opens Step 5 Preview API at $1 Per Million Input Tokens with a 600B Sparse MoE Architecture

StepFun announced Step 5 Preview on September 20, 2026, and opened API access the same day at $1.00 per million cache-miss input tokens, per Pandaily. The model is a 600B-parameter sparse mixture-of-experts with approximately 27B parameters active per token and a 1 million-token context window. StepFun targets it at long-horizon agent workloads including AI coding, software engineering, financial analysis, and professional knowledge tasks. Cache hits cost $0.05 per million; output costs $2.70 per million. Per StepFun's analysis, the model scores 44 on Artificial Analysis's intelligence index, matching Kimi K3 Max and running at roughly one-seventh the cost of GPT-5.6 Sol at the same score level. Open weights are scheduled for October 15; the Hugging Face repo currently shows only a placeholder. Operators evaluating long-context coding agents at the $1/million price point have a real API to test before the weight release.

Full story: StepFun Step 5 Preview 600B MoE and open weights timeline

California Governor Newsom Orders Two-Month Working Group to Evaluate an AI Kill Switch Mandate

Governor Gavin Newsom signed an executive order on September 18, 2026, directing a state working group to publish recommendations within two months on AI safety measures, per the California Governor's Office. Two requirements are on the table for evaluation: a mandatory emergency "kill switch" capable of halting frontier model operations, and a mandate for independent third parties to draft and monitor safety plans at AI companies operating in California. Newsom called on the federal government to act "before it's too late." The order itself does not mandate either measure; it sets a recommendation deadline of approximately November 18, with legislation or further executive action expected to follow. AI operators with California-based deployments or offices should track the working group's November output. Whether New York or Texas respond with parallel orders is a secondary signal.

Full story: Newsom AI kill switch executive order and two-month review

Sources