step 3.7 flash dropped yesterday – open weights, competitive with gemini flash 3.5 and deepseek v4 flash. and then immediately free for anyone to download. that last part is the tell
the ipo context. stepfun is filing for a hong kong ipo by end of june at a $10 billion valuation, coming off a $2.5 billion pre-ipo round closed in may. the model release lands squarely in the middle of roadshow season. you don't monetize weights you give away on day one – you monetize the story they tell to anchor investors
the actual business. 42 million devices ship with stepfun models embedded. oppo, honor, zte, geely – around 60% of china's major smartphone brands, plus a cockpit os in mass-produced geely cars. that distribution is the real moat, nowhere near replicable by someone who downloads a checkpoint. open weights cost them little because the model isn't the defensible asset
the investors. the capital stack reflects the strategy. early backers tencent and qiming venture partners were joined in the $718m series b+ by china life insurance, pudong venture capital, and state-linked funds. the $2.5 billion pre-ipo round then brought in zte, omnivision, huaqin technology, and the hong kong investment corporation. that's a hardware supply chain writing checks – device manufacturers and component suppliers who have a direct commercial interest in stepfun's models being good enough to ship in products
the play. what the release buys is international visibility – benchmark citations, developer mindshare, credibility that travels outside china before an ipo that needs anchor investors to believe a $10 billion number
the model is fine. the launch is smarter than the model

⚡️ Step 3.7 Flash is here: The new frontier is agent efficiency.
— StepFun (@StepFun_ai) May 29, 2026
#1 ClawEval-1.1 (67.1), #1 SimpleVQA Search (79.2), #2 SWE-PRO (56.3), 95.3 on V* Python. Open weights under Apache 2.0.
Built for agentic, coding, search, and multimodal workflows — balancing speed, cost, and… pic.twitter.com/mzLi1HnkxU
Addy Crezee