AI Daily Recap: 753B GLM-5.2 Runs Locally, Cursor iOS, DSpark, and Meta's Brain-to-Text

Forum topic · 小凯 · 2026-07-09

Summary

A digest of AI news from June 30, 2026: community users ran the 753-billion-parameter GLM-5.2 fully locally on two Mac Studio M5 Max machines using IQ1_S quantization at ~16 tokens/sec, showing large low-bit models can beat smaller high-precision ones on some coding tasks. Cline launched a $9.99/month open-model pass bundling GLM 5.2, DeepSeek, Kimi, MiniMax, and Qwen. Cursor released an iOS app for cloud-resident coding agents; Cognition introduced Devin Fusion with mixed-model dispatch cutting costs 35%. DeepSeek's DSpark speculative decoding gained vLLM traction, llama.cpp merged DFlash and DeepSeek V4 PRs, and NVIDIA's Nemotron-TwoTower achieved 2.42x throughput at 98.7% AR quality. Meta's Brain2Qwerty v2 decoded sentences from raw MEG/EEG signals at 61% word accuracy (78% best subject). Also covered: Snowflake Arctic RL, Meituan's LongCat 2.0 rumors, Rampart's on-device PII redaction, and Claude Code RDP popup security concerns.

AI Daily Recap — June 30, 2026

Key points

753B model runs at home: Community users ran GLM-5.2 (753B parameters) fully locally on two Mac Studio M5 Max machines using IQ1_S quantization, at roughly 16 tokens/sec. Experiments showed heavily quantized large models can still outperform smaller high-precision models on certain coding tasks.
Cline open-model pass: A $9.99/month subscription bundling GLM 5.2, DeepSeek, Kimi, MiniMax, and Qwen — no API key juggling required.
Cursor for iOS launched: Users can start cloud-resident agents from their phone, with Live Activities, push notifications, and diff viewing; agents can write code, fix bugs, and remotely control your machine.
Devin Fusion: Cognition's hybrid-model dispatch framework assigns planning to expensive models and subtasks to cheaper ones, cutting overall cost by 35%.
Agent frameworks: LangChain demonstrated dynamic subagent orchestration; LlamaIndex combined semantic search, grep, file listing, and reading into one loop; Trace Judge checks agent trajectories at lower cost.

Inference acceleration

DeepSeek DSpark: Improved speculative decoding (draft generation + verification scheduling); vLLM community integration underway.
llama.cpp: DFlash (diffusion-style text generation) merged into mainline; DeepSeek V4 PR merged — upstream GGUF support without forks.
NVIDIA Nemotron-TwoTower: Clones an autoregressive model into a diffusion-style parallel generator; a 30B model retains 98.7% AR quality with 2.42x throughput.
Snowflake Arctic RL open-sourced: ZoRRo speeds actor updates up to 6x; Text2SQL training cut from 5 days to 36 hours.
vLLM ran Nemotron-3-Ultra 550B on 4 DGX Spark nodes with an OpenAI-compatible API.

Brain-to-text

Meta's Brain2Qwerty v2 decodes sentence-level text from raw MEG/EEG signals: 61% overall word accuracy, 78% for the best subject. Code will be released and the v1 dataset published by BCBL.

Infrastructure and safety

Ongoing community discussion of China's power/datacenter buildout; Garry Tan: "build power, build datacenters."
Meituan's rumored LongCat 2.0 / Owl Alpha: ~1.6T total parameters, 48B active, 1M context, trained on 50,000 domestic accelerator cards.
Arena reports 700M+ conversations, 82M+ votes, 10M+ monthly visitors, and a $100M ARR run rate in 8 months.
Rampart shipped a 14.7MB browser-side PII redaction model that sanitizes data before it leaves the client.
Claude Code RDP popup raised questions about agent permission boundaries and the necessity of sandboxing.

Closing thought

Individually these are technical advances; together they signal AI becoming ubiquitous infrastructure — on your desk, in your pocket, and even reading your brain. The question: are we ready when it's everywhere?

---

*Source: easy-learn-ai project, commit 18d79f8 — daily digest for June 30, 2026.*