Daily Brief
2026-08-21 · Evening · 8 items
-
Hugging Face releases LFM2.5 DSpark draft models, up to 3.18× GPU throughput
Around 300M parameters cut function-call latency by 57% on average with no quality loss.Hugging Face：Blog（RSS）
-
Mistral Launches Agentic Search for Multi-Step Retrieval
Five-tool retrieval loop turns one-shot RAG into a verified multi-step workflow.Mistral AI：News（网页）
-
Alibaba launches Qwen-UI-Agent to make models truly use every screen
It covers mobile, PC, web, and DeepSearch environments—Alibaba's first general-purpose GUI agent base model.IT之家（RSS）
-
OpenAI CFO tells staff IPO will close by 2027 at the latestRECAP
Annualized revenue grew 35% this quarter, with enterprise up 50%, and a debut could come sooner.IT之家（RSS）
-
OpenRouter announces merger with StripeRECAP
It handles 10T+ tokens daily across 400+ models, and will keep its name and roadmap post-merger.OpenRouter：Announcements（RSS）
-
LMSYS optimizes DeepSeek-V4-Pro on H20, closing in on B300RECAP
The 1.6T-param MoE hits 271 tokens/s on a single H20-141GB, just 1.42x behind B300.LMSYS：Blog（Chatbot Arena 团队）
-
GLM-5.3 API launches: scores 60 on AA index, ties for first among open-source modelsRECAP
Lowest per-task cost among frontier flagships, pricing unchanged from 5.2, weights open-sourced next Friday.公众号：智谱（GLM）
-
Claude designs protein binders: 26.7% hit rate, roughly double the industry normRECAP
Mythos Preview hit 26.7% across 15 targets, versus the typical 10–15% industry range.Anthropic：Research（发表成果 · 网页）