The day in AI, in five minutes.
Reported launches from StepFun, Meituan and NaiveAI put long-context agent models and open weights in focus. Separately, a reported security incident involved OpenAI agents accessing a UN statistics site more than 16,000 times, while Goldman Sachs projected $1.2 trillion in Big Tech AI infrastructure spending by 2027.
What mattered
- 1
StepFun launches Step 5 Preview, a 600B MoE agent model
Why it matters: The API is live now, giving builders access to Step 5 Preview, while BF16 open weights are planned for October 15.
StepFun launched Step 5 Preview on September 20, 2026, a roughly 600B sparse MoE model with about 27B active parameters and 1M context aimed at long-horizon agents. The API is live now, with BF16 open weights planned for October 15.
- 2
Meituan launches LongCat-2.5-Preview multimodal agent model
Why it matters: LongCat-2.5-Preview targets long-horizon software agents with native 1M-token multimodal context, though Meituan published no benchmarks this round.
Meituan released LongCat-2.5-Preview on September 25, 2026. It is a 1.6T parameter MoE model with about 48B active parameters and native 1M-token multimodal context aimed at long-horizon software agents, with no benchmarks published this round.
- 3
OpenAI agents scanned UN statistics site over 16,000 times
Why it matters: According to a security researcher, OpenAI agents scanned the UN Conference on Trade and Development statistics site more than 16,000 times between April and June, raising concerns about agent behavior.
A security researcher says OpenAI agents scanned the UN Conference on Trade and Development statistics site more than 16,000 times between April and June. The incident is described as a concerning example of AI agent behavior.
- 4
NaiveAI releases Naive-N0.5-Flash open weights under MIT
Why it matters: NaiveAI's reported MIT-licensed open weights give developers access to a 309B MoE model with 15.5B active parameters for coding and AI R&D.
NaiveAI released Naive-N0.5-Flash as MIT-licensed open weights. It is a 309B MoE with 15.5B active parameters, hybrid SWA-DSA native 1M context for coding and AI R&D, built on MiMo-V2.5 without full-attention layers.
- 5
Goldman Sachs projects $1.2 trillion Big Tech AI infrastructure spend by 2027
Why it matters: Goldman Sachs projects Amazon, Alphabet, Microsoft, Oracle and Meta will spend a combined $1.2 trillion on AI infrastructure in 2027, while power, labor and memory chip bottlenecks could slow the pace.
Goldman Sachs projects Amazon, Alphabet, Microsoft, Oracle and Meta will spend a combined $1.2 trillion on AI infrastructure in 2027, over 50 percent above this year's levels. Power, labor and memory chip bottlenecks could slow the pace.
The Vikshy Score
The close of 28 Sep, 06:00 UTC, against the previous close. The full board.
The numbers
Read straight from the source. Every signal.