Daily AI brief · September 20, 2026
StepFun Step 5 Preview Lands at $1 per Million
StepFun's 600B-parameter Step 5 Preview arrived with open weights promised for October, as Qwen cut audio pricing and tightened an image license and a 27B model shrank to fit a laptop.
StepFun launches Step 5 Preview, with open weights due October 15 (MarkTechPost, Pandaily) Step 5 Preview is a sparse mixture-of-experts model with 600 billion total parameters, 27 billion active per token and a 1-million-token context window, taking text, image and video input. It launched as a paid API on September 20 at about $1 per million input tokens and $2.70 per million output, and scores about 44 on the Artificial Analysis intelligence index. StepFun is aiming it at long-running agent work like coding and financial analysis. What it means for SMBs / MSPs: Another capable model at a low per-token price is useful leverage, but wait for the October weights and independent tests before you move client work onto it.
Qwen3.8-Omni-Flash cuts audio input costs by more than 98% (MarkTechPost, Startup Fortune) Alibaba's Qwen team released Qwen3.8-Omni-Flash on September 18. It takes text, images, audio and video, returns text, and has a 1-million-token context. QwenCloud lists it at $0.15 per million input tokens and $0.47 per million output, and audio input per hour is more than 98% cheaper than before. Unlike many past Qwen releases, it's API-only with no open weights. What it means for SMBs / MSPs: Voice and video agents such as call summaries or meeting review just got much cheaper to run, but check where the data is processed before you send it client calls.
Qwen-Image-2.1 drops Apache 2.0 for a research-only license (Qwen on Hugging Face, eesel AI) The new 7B image generation and editing model, released September 20, adds transparent backgrounds and accepts up to ten reference images. The weights are downloadable, but the Qwen Research License bars commercial use without a separate agreement. That's a break from earlier Qwen-Image releases, which stay under Apache 2.0. What it means for SMBs / MSPs: "Open weights" doesn't mean "free to use in business," so read the license file before a client ships anything made with it.
PrismML squeezes a 27B model into 5.9 GB (MarkTechPost, PrismML) Ternary Bonsai 2 27B rewrites Qwen3.8 27B's weights as -1, 0 or +1, cutting memory use by more than nine times. PrismML reports it keeps 98.2% of the original's score across 20 benchmarks, and it's released under Apache 2.0 for commercial use. What it means for SMBs / MSPs: Useful local models now fit on ordinary hardware, which makes an on-premises option realistic for clients who can't send documents to the cloud.
Ternary Bonsai 2 27B on a spare machine. This is for MSPs and champions with a client that has sensitive documents and no appetite for cloud AI. A first test takes about an hour. First step: download the model from PrismML's release page, run it on a workstation with a recent GPU, and ask it the same five questions about internal documents that you'd ask a cloud model. Compare the answers and note where it falls short.
Pull last month's AI bills and list what each subscription and API key is for. Model prices keep falling: Qwen3.8-Omni-Flash cut audio input costs by more than 98%, and Step 5 Preview launched at about $1 per million input tokens. If you signed up at last year's rates and never looked again, you may be overpaying. Pick one change to test, and read the license and data terms before moving any work.
*For MSPs:* A one-time AI spend review that shows each workload's cost on a cheaper current model produces savings clients see on the next invoice.
Picking a model on a leaderboard number and a launch price. Step 5 is a preview and its weights aren't out yet. Qwen's newest image model changed its license terms. Run your own test on your own tasks and read the terms before committing a client.
Open models are splitting into "open to download" and "open to use." Some labs keep shipping permissive licenses while others hold back weights or restrict commercial use. For small businesses, the license is now as important a buying criterion as the benchmark score.
Compiled on September 27, 2026 from reporting published on or before September 20, 2026.