Daily Brief
2026-08-04 · Evening · 12 items
-
Swiftlet runs 80B Qwen on a Mac with just 4.3GB of RAM
A Swift+Metal runtime streams expert weights on demand — even an iPhone can run the 35B version.Hacker News 热门（buzzing.cc 中文翻译）
-
Cloudflare launches Agents platform with unified session view
All agent sessions on the platform in one view, plus first-of-its-kind agent tracing.Cloudflare Blog
-
Cloudflare's AI software factory nearly zeroes out Astro's issues
Isolated AI sub-agents reproduce, diagnose and fix bugs — open issues fell from 200+ to about 30.Cloudflare Blog
-
Cloudflare lets agents debug Workers with local tracing
wrangler dev now auto-captures OpenTelemetry traces — no SDK needed, agents can read them directly.Cloudflare Blog
-
Cloudflare ships CI SDK for CI/CD at millions-of-repos scale
Built on Workflows and Sandbox SDK — builds, tests, dependency caching and conditional deploys.Cloudflare Blog
-
Cloudflare proposes ADLC: agents take over the dev lifecycle
An agent development lifecycle to replace SDLC — agents extend from writing code to running more of the process.Cloudflare Blog
-
DeepSeek V4 Flash runs in production on a single AMD MI300X
An open-source repo ships full configs and patches — the 304B model runs unquantized on 192GB HBM.Hacker News 热门（buzzing.cc 中文翻译）
-
China issues first mandatory national standard for L3/L4 self-driving
The safety requirements for automated driving systems take effect on July 1, 2027.IT之家（RSS）
-
Confirmed: the 80% GPT-5.6 Luna price cut is permanent
The cut is not a stunt — efficiency gains mean the lower price is here to stay.X：Tibo (@thsottiaux)
-
MiniCPM open-sources ForgeStencil: AI auto-optimizes 100+ apps in a week
Kernel and App agents work in a closed loop with zero human input — auto research, optimize and deploy.公众号：面壁智能（MiniCPM）
-
SwanTale: one model for multi-speaker speech and audio generation
Unifies zero-shot and instruction tasks, with a data recipe tackling speech-data scarcity.HuggingFace Daily Papers（社区热门论文）
-
UEmbed: sparse and dense multimodal embeddings in one pass
A decoder-only model yields word-level sparse and dense representations in a single causal pass.HuggingFace Daily Papers（社区热门论文）