EAIDaily — September 18, 2026

English AI News Brief | Focus: AI Coding & Embodied Intelligence

1. Anthropic Relaunches Claude Code Projects as a Multi-Agent Cloud Orchestrator

What happened: On September 17, Anthropic rolled out a redesigned Projects feature inside Claude Code. Instead of a single chat thread, a Project now runs multiple cloud-based agent “threads” in parallel, each on its own git branch and repo copy, coordinated by a top-level agent. Threads can spawn subagents and structured workflows, with shared memory, goals, and a project library. Merge conflicts between threads are surfaced and resolved like normal pull requests. Access starts with select Pro/Max subscribers; local-tool support is promised “very soon.”

Why it matters: This is Anthropic’s clearest move to turn Claude Code from a pair-programmer into a project-level execution layer. The bet is that the bottleneck in AI-assisted engineering is no longer raw code generation but coordination — running many agents in parallel without them stepping on each other. If local support ships on time, it directly challenges Cursor, Grok Bot, and other multi-agent coding stacks.

Sources: The Verge; Anthropic; AI Chat Daily; Brocker

2. Z.ai Builds GLM-5.3-Flash Inference Stack on 100,000+ Chinese Chips — with an Infra Agent

What happened: Z.ai published a technical account on September 17 describing how it built a production-grade inference service for GLM-5.3-Flash from scratch on a cluster of more than 100,000 Chinese-made AI accelerators. Much of the optimization work was carried out by an Infra Agent powered by GLM-5.3 itself. The team reports end-to-end throughput improved roughly 3.2× from first run to production readiness in about two weeks. Key techniques included ReplaySSM, tensor parallelism, mixed INT8/FP8/BF16 cache quantization, and an Encode-Prefill-Decode disaggregated architecture. The company says all production inference for GLM-5.3-Flash now runs on this cluster.

Why it matters: This is one of the first credible cases of a frontier model helping optimize the infrastructure that serves it — an early, narrow form of recursive self-improvement. It also demonstrates that Chinese AI labs are closing the gap on domestic-hardware deployment at massive scale, reducing reliance on restricted Nvidia chips for inference serving.

Sources: Z.ai blog; Unite.AI; 163.com; ExplainX.ai

3. Xiaomi MiMo-V2.6 Goes Live: Public RL Training Dashboard for a 1T-Class Model

What happened: Xiaomi’s MiMo team, led by former DeepSeek researcher Luo Fuli, launched a live public dashboard on September 17 streaming the reinforcement-learning training of MiMo-V2.6-Pro and MiMo-V2.6-Flash in real time. The dashboard shows reward curves, token throughput, step timings, environment mix, and running cost. As of mid-September 17, the two runs had consumed more than $1.2 million in compute, processed tens of billions of tokens, and reached 63.72% on DeepSWE v1.1 for Pro — a large jump from the V2.5 baseline. Code-heavy tasks make up about two-thirds of the training mix.

Why it matters: Xiaomi is turning training transparency into a competitive signal against the closed-door frontier-lab model. The dashboard also gives a rare public view of large-scale agentic RL — long-horizon episodes, asynchronous rollouts, and rubric-based reward assignment — which is becoming the central recipe for coding-agent capability gains.

Sources: mimo.xiaomi.com/rl; Forkast News; 36Kr; Sina Finance

4. Google Home MCP Opens the Smart Home to Any AI Agent

What happened: Google opened Home MCP early access on September 16. It is an official Model Context Protocol server that lets any MCP-compatible agent — Google names Claude, ChatGPT, Antigravity, Hermes, and OpenClaw — enumerate devices, read live states, query event history, and run actions across the Google Home ecosystem. A separate Home Developer MCP server exposes Home API, Matter, and Thread documentation to coding agents. Access is currently limited to Google Home Premium Advanced subscribers ($20/month) in the U.S.

Why it matters: This is the first time a major smart-home ecosystem has been opened to third-party AI agents through a standardized protocol. It signals a shift from brand-locked voice assistants to an agentic layer that can reason across devices, history, and physical context. The line between software agents and the physical home is getting thinner — a direct bridge to embodied intelligence use cases.

Sources: Google Home developers; Unite.AI; AI/TLDR; The Verge; TMTPost

5. Huawei Unveils Ascend 960DT/960PR Roadmap, Targets Annual Chip Generations

What happened: At Huawei Connect 2026 on September 17, rotating chairman Wang Tao announced the Ascend 960DT (training) will launch in Q1 2027 and the Ascend 960PR (inference) in Q3 2027, ahead of the previous schedule. Huawei also committed to annual generational upgrades through Ascend 970 (2028) and Ascend 980 (2029), with each generation roughly doubling compute. To offset a still-narrow single-chip gap versus Nvidia, Huawei is scaling to 4,096-card supernodes and says a 100,000-card cluster built around such supernodes can achieve ~2.75× better model-FLOPs utilization than conventional 8-card-server clusters.

Why it matters: The roadmap is a concrete milestone in China’s push for AI compute sovereignty. If the timeline holds, Huawei gives domestic labs a credible alternative for both training and inference at frontier scale — directly relevant to the cost and availability of the compute that powers both coding agents and embodied-AI models.

Sources: Caijing Magazine; PixelsTech; The Insight Asia; Industry Wired

6. Songyan Dynamics Launches Scalabot Embodied-Intelligence Brand and HERON Model Stack

What happened: Beijing-based humanoid maker Songyan Dynamics unveiled its new embodied-AI sub-brand Scalabot on September 8, followed by two HERON-architecture models: HERON-WorldModel (a world model trained on a data pyramid of real robot data, reconstructed simulations, and first-person human videos) and HERON-CRA (Context Reinforcement Action Model) on September 15. Founder Jiang Zheyuan says the company has shifted from pure hardware to “brain” development now that its bodies are reliable and low-cost enough for scaled deployment. Songyan plans to put hundreds to thousands of home robots into real households to collect data and close the data flywheel.

Why it matters: Songyan’s pivot reflects a broader industry shift: hardware bodies are no longer the bottleneck; data and intelligence are. The company’s “four red lines” — avoid over-reliance on real-robot data, avoid single-body binding, avoid simulation-only training, and avoid demo-driven development — are a concise summary of the failure modes currently holding back embodied AI.

Sources: 163.com; Toutiao/Everyday Economic News; Beijing.gov.cn; Caijing NetEase

7. Edge0 Runs a 35B MoE Coding-Capable Model from an SSD in 2.9 GB of RAM

What happened: The open-source Edge0 framework, released September 8, shows how a 35B-parameter Mixture-of-Experts model can run on a laptop with peak active memory of just 2.9 GB. It keeps most expert weights on SSD and uses a “prerouter” to prefetch only the experts the next token will need. The flagship edge0-35b checkpoint is built on Qwen3.5-MoE 35B-A3B, runs on Apple Silicon via MLX, and loses only about 4 points on average across AIME 2026, HumanEval, GPQA-Diamond, MMLU-Pro, and IFBench compared with the FP16 base.

Why it matters: Edge computing for coding agents has been limited by memory, not model quality. By making a 35B-class model runnable on ordinary laptops, Edge0 expands the addressable surface for private, local coding agents and lowers the hardware barrier for developers who cannot rent GPUs or send code to cloud APIs.

Sources: Edge0-AI/Edge0 GitHub; AI/TLDR; MindPattern; QuidProQuo model card; Geeky Gadgets

8. ScienceIDE Turns Scientific Code Repositories into Agent Training Environments

What happened: A 45-researcher team led by PhAI Labs released ScienceIDE (arXiv:2609.19134) on September 16. It converts real scientific codebases — covering astrophysics, ocean biogeochemistry, plasma kinetics, quantum many-body physics, and more — into programmable, verifiable training environments for scientific agents. Tasks include repairing injected semantic defects and reimplementing excised routines. Reward comes from “the simulation is numerically correct again,” not from diff matching. Models trained on these trajectories — PhAI-IDE-4B/9B/72B — show positive transfer to general coding, reasoning, and knowledge benchmarks.

Why it matters: ScienceIDE attacks a core problem for coding and research agents: verifiable reward in messy, real-world codebases. If scientific repositories can be turned into repeatable gyms with physics-based judges, they become a massive, underutilized source of high-quality training signal for agentic coding models.

Sources: arXiv 2609.19134; GitHub aitofound/ScienceIDE; AI Modeling; Clauday

Thematic Takeaways

Coordination is the new code-generation. Multi-agent orchestration (Claude Code Projects) and transparent large-scale RL (MiMo-V2.6) show the field moving past single-turn coding into fleet management and long-horizon training.
Agents are eating infrastructure. Z.ai’s Infra Agent and ScienceIDE’s verifiable scientific environments both point to the same pattern: the most valuable work is now building the environments and feedback loops that let agents improve systems — including the systems that run agents.
Physical-world boundaries are thinning. Google Home MCP extends software agents into the home, while Songyan Dynamics and Huawei’s chip roadmap show embodied AI converging on data flywheels and domestic compute stacks.
Local and sovereign deployment are accelerating. Edge0’s SSD-offloaded 35B model and Huawei’s Ascend 960 roadmap both lower dependence on centralized, foreign-controlled compute and APIs.

Compiled by WorkBuddy automation | Focus: AI Coding + Embodied Intelligence | 8 stories selected