Daily Brief
2026-09-22 · Evening · 8 items
-
StepFun Step 5 Preview tested by Artificial Analysis
Scores 44 on the Intelligence Index, tied with Kimi K3 (max), at roughly $0.7 per task.X：Artificial Analysis (@ArtificialAnlys)
-
OpenRouter breaks down NVIDIA's Nemotron 3.5 Lightning
A 30B-parameter MoE that activates only ~3B, tuned for tool calls and coding in agent loops.OpenRouter：Announcements（RSS）
-
Tencent Hunyuan releases Hy Image3.5 preview image model
Blind tests show over 30% higher win rate vs. the prior version, with up to 2K output.腾讯混元：Research（API）
-
Kazzik's hands-on: Grok 4.7 underperforms, MiMo V2.6 leads overall
Across performance, price, and inference speed, MiMo V2.6 beat Grok 4.7 in all three.公众号：数字生命卡兹克
-
xAI launches Grok 4.7, pitched as its strongest model for coding and knowledge work
Priced at $2 per million input tokens and $6 per million output tokens, same as 4.6, with a faster variant that doubles both speed and cost.xAI：News（网页）
-
Nathan Lambert maps the US-China open-source AI gap
Chinese models have outpaced US counterparts by 1.6B downloads since July 2025, doubling the US total.Nathan Lambert：Interconnects（RSS）
-
Xiaomi's MiMo-V2.6-Pro reaches ~10th on Code Arena WebDevRECAP
It ranks around 3rd among open-weight models, scoring 153 points higher than MiMo-V2.5-Pro.X：Arena (@arena)
-
Xiaomi Releases MiMo-V2.6 Pro and Flash, Open-Source Omnimodal ModelsRECAP
Pro scores 46 on the Artificial Analysis Intelligence Index, the highest among open-source models.X：小米 MiMo (@XiaomiMiMo)