EAIDaily — 2026-07-28

Daily AI news brief focused on AI Coding and Embodied Intelligence, curated for English-speaking builders tracking frontier AI developments. Compiled 2026-07-28 from AI HOT, vendor blogs, and tech-press search results.

At a Glance

Two themes dominate the news cycle. First, the wires underneath coding agents just got rewritten in production: MCP’s largest revision ever ships on July 28 as a stateless protocol, GitHub already runs its MCP server behind a round-robin load balancer with no Redis, and Kimi K3’s open weights drop its full infra stack (MoonEP / FlashKDA / AgentENV) on the same day Anthropic clarifies it never opposed open-weight models. Second, embodied AI is leaving the “demo era” and entering the “long-horizon + delivery” era: τ₀-VLA brings test-time compute and world-model reasoning to robot sub-task planning, TARS wins WAIC’s SAIL award with a 1-million-hour industrial foundation model, and Tesla concedes Optimus still has three unsolved problems while Chinese makers ship ~90% of humanoids shipped globally. Today’s eight items cluster naturally around those two shifts.

Headline Items

1. MCP 2026-07-28 Spec Ships as the Protocol’s Largest Revision Ever

Category: AI Coding / Agent Infrastructure Source: Model Context Protocol blog, The Register, bruno.digital Date: 2026-07-28

The final MCP specification for 2026-07-28 publishes today. The protocol is now stateless at the wire level: the initialize handshake and Mcp-Session-Id header are removed, protocol version and client info travel in _meta on every request, and any MCP server can sit behind a plain round-robin load balancer. Three smaller changes enable routing and ops at scale: mandatory Mcp-Method and Mcp-Name HTTP headers for gateway-level routing, ttlMs / cacheScope cache hints modeled on HTTP Cache-Control, and W3C Trace Context propagation in _meta. Long-lived SSE streams are replaced by Multi-Round-Trip Requests, and two experimental features graduate into a formal Extensions framework: MCP Apps (server-rendered HTML UI in sandboxed iframes) and the Tasks extension (long-running work via tool-returned handles). Authorization gets six OAuth/OIDC hardening changes including RFC 9207 issuer validation.

GitHub shipped same-day support on July 23, removed Redis-backed session storage and stopped inspecting request payloads at the gateway. The spec also introduces a formal 12-month deprecation policy for Roots / Sampling / Logging.

Why it matters: This is the inflection point where MCP stops being a research protocol and becomes infrastructure you can standardize a fleet on. Every team that held off on remote MCP because of sticky-session complexity now has a clear migration path. The deprecation clock gives platform engineering a calendar to plan against — the first time the protocol has committed to a predictable breaking-change lifecycle since launch.

2. Kimi K3 Open-Sources Weights and Three Training-Infra Layers on the Same Day

Category: AI Coding / Open-Weight Frontier Models Source: Moonshot AI (X / blog), LMSYS, Hugging Face, GitHub Date: 2026-07-27

Moonshot AI released Kimi K3 — a 2.8-trillion-parameter Mixture-of-Experts model with native visual understanding and a 1-million-token context — under a Modified MIT license. The release ships alongside three open-source infrastructure libraries that made the model possible: MoonEP (a fine-grained MoE communication library using online integer linear programming and zero-copy transfers to keep expert-parallel efficiency under load imbalance), FlashKDA (Kimi Delta Attention kernel running 1.72–2.22× faster prefill than the flash-linear-attention baseline on NVIDIA H20), and AgentENV (a sandbox system co-developed with KVCache.ai that supports fast snapshotting, restoration, and forking for massively parallel agent training).

SGLang and Miles added Day-0 support: single-card batch-1 decoding reaches ~113 tokens/sec, and ~423 tokens/sec with DSpark speculative decoding. Modal trained a custom DFlash speculator for K3 that ships at Day-0. Reportedly, ~3,700 developers queued on Hugging Face’s preload page before the upload went live.

Why it matters: K3 is the first self-hostable frontier-class coding model in the 3T-class. Together with the open infra stack, teams can now reproduce Moonshot’s training playbook instead of reinventing expert parallelism, hybrid-attention kernels, and agent sandboxes. The price pressure on closed-model APIs is structural: K3 reaches Fable 5 / GPT-5.6 Sol-tier performance and the weights are free.

3. τ₀-VLA Brings Test-Time Compute and World-Model Reasoning to Robot Long-Horizon Tasks

Category: Embodied Intelligence / Foundation Models Source: Shanghai Innovation Institute (SII), AgiBot Robotics, GitHub sii-research/tau-0-vla Date: 2026-07-27

Shanghai Innovation Institute (Luo Jianlan team) and AgiBot Robotics jointly released τ₀-VLA, a hierarchical Vision-Language-Action foundation model for long-horizon manipulation. The architecture decouples “slow thinking” (high-level subtask planning) from “fast execution” (low-level control), introduces Test-Time Scaling and world-model-guided reasoning into embodied decision-making, and was pretrained on 40,115 hours of heterogeneous real-world robot data — including more than 20,000 hours of real-machine trajectories across multiple embodiments.

A Qwen3.5 vision-language backbone is paired with a Mixture-of-Transformers action expert trained through conditional flow matching on a unified 40-D state/action interface. End-to-end success on a Milk Tea sequence rose from 5/10 to 7/10 with test-time compute; book tidying from 6/10 to 9/10; room cleaning from 5/10 to 7/10. Next-subtask prediction accuracy: 87.3% (Milk Tea), 88.0% (Book Tidy in-distribution), 74.0% (Book Tidy out-of-distribution), 87.0% (Room Cleaning). Model, paper, and code are released under Apache 2.0.

Why it matters: τ₀-VLA answers the question τ₀-WM could not: even when a low-level VLA generates correct actions, it still cannot judge whether those actions match the correct stage of a long-horizon task. By lifting test-time compute out of the action path and into the subtask boundary, the model demonstrates that robot intelligence scales not only with parameters and data but with reasoning budget allocation — the same insight that gave us o1-style reasoning in text models.

4. TARS AWE 3.5 Wins WAIC 2026 SAIL Award, Validates Million-Hour Industrial Foundation Model

Category: Embodied Intelligence / Foundation Models Source: TARS (PRNewswire / EQS), WAIC 2026 Date: 2026-07-23

TARS debuted AWE 3.5 (AI World Engine), its latest embodied-native foundation model, at WAIC 2026 and won the SAIL Award (Superior AI Leader) for technical innovation and industrial potential. AWE 3.5 is trained on 1+ million hours of human-centric real-world data and integrates action, perception, geometry, and tactile sensing in a unified framework; TARS claims ~2× task-execution efficiency vs. Pi 0.5 with closed-loop behavior across multi-minute tasks. A first-of-its-kind “pre-training + post-training” paradigm is designed for reproducible, scalable iteration.

Live demos included phone packing, backpack organization, precision screw sorting, and a full-scale circular automotive wire-harness production line using multiple A1 robots (selected for WAIC’s “Model Era · Partner City” Smart Manufacturing Hub showcase). DexHand, mounted on an A1, performed live card spreading, handwriting, and Rubik’s Cube solving alongside magician Deng Nanzi. TARS is working with Shanghai’s Jiading District to validate the technology at the thousand-unit industrial cluster scale and plans to expand the pre-training dataset to 10 million hours by end of 2026.

Why it matters: AWE 3.5 is the first embodied foundation model from a Chinese startup with public SAIL recognition and an explicit roadmap to million-unit industrial clusters. It is the clearest signal yet that embodied AI is splitting into a “data-flywheel” camp (TARS, Xiaomi-Robotics-1, Tencent Apexio) and a “fleet-deployment” camp (Zhiyuan, Yushu, AgiBot), with foundation-model capability deciding which camp dominates the long run.

5. Anthropic Clarifies It Never Advocated Banning Open-Weight Models; Pushes for Compute Export Controls and Mandatory Safety Testing Instead

Category: AI Coding / Policy & Open-Weight Strategy Source: Anthropic Newsroom, 财联社 (CLS), AI HOT Date: 2026-07-27

Anthropic CEO Dario Amodei published the company’s formal position on open-weight models: Anthropic has never advocated banning open weights, supports chip export controls to China, supports cracking down on industrial-scale knowledge distillation, and supports mandatory safety testing for all sufficiently-capable models. He argues that “protectionist bans” cannot solve the actual national-security concerns — authoritarian governments using more capable AI for military advantage or surveillance.

The clarification matters because it came a day after NVIDIA, Microsoft, Hugging Face, IBM, Dell, SpaceX, Adobe, and CrowdStrike launched the Open Secure AI Alliance (OSAA) on July 27 — a coalition to develop open AI safety and cybersecurity tooling on top of open models. Jensen Huang’s first X post was a re-share of an open-weight defense letter signed by 20+ organizations (Musk, Satya Nadella, etc.). OpenAI and Anthropic declined to sign; Sam Altman publicly said he hopes “America wins on open AI.”

Why it matters: The fault line in 2026 frontier AI is no longer “open vs closed” philosophy — it is converging into “what level of openness can be combined with national-security controls and verifiable safety testing.” Two parallel consensuses are forming simultaneously: a pro-open-weight coalition (NVIDIA + Meta + MS + IBM + Y Combinator ranks) and an export-control + safety-test policy consensus inside Anthropic and parts of Washington. Both are policies for the same future of openly-distributed intelligence.

6. GitHub Copilot Launches the “Harness” Workflow: Prototype, Plan, Implement, and Code Review in One Tool

Category: AI Coding / Agent UX Source: GitHub Blog Date: 2026-07-27

GitHub announced the Copilot “Harness” workflow, an opinionated end-to-end pipeline that keeps developers inside one AI tool from prototyping → planning → implementation → code review, instead of stitching together Cursor / Codex / Claude Code / Aider / Continue / terminal agents. The launch post frames it as “the harness is all you need (mostly)” — an explicit pushback on tool fragmentation.

Alongside it, the GitHub Copilot app moved from preview to a multi-agent workspace, with /create-canvas for in-browser UI preview and Agent Merge handling PR review feedback and merge conflict resolution autonomously.

Why it matters: Coding-agent fragmentation has been the norm for two years. The Harness is GitHub’s bet that the next interface layer is the agent harness — the glue that selects tools, manages subagents, and handles task lifecycle — not the model. If the bet pays off, developer attention consolidates around a small number of opinionated harnesses and the underlying model APIs become interchangeable commodities.

7. Ten Million Parameter Embodied Foundation Models Shift Industry Center of Gravity Toward Data and Edges

Category: Embodied Intelligence / Trend Synthesis Source: CD RCWAPP analysis, AI HOT daily, TARS PR, Tencent WAIC lineup Date: 2026-07-23 to 2026-07-28

Multiple independent releases this week crystallized the same trajectory: embodied foundation models (EFMs) are racing past 100B parameters, real-trajectory data is the new moat, and on-device NPU is the new bottleneck. NVIDIA GR00T N2, Alibaba Lingxi, and Google RT-X 3 are all above 100B in 2026; Figure founder Brett Adcock has publicly committed to a trillion-parameter model in 2027. Open X-Embodiment has grown from 5M to 200M tracks spanning 200+ robot types and 50,000+ tasks; Xiaomi-Robotics-1 alone ingested 100,000 hours across 1,700+ scenario types. TARS plans 10 million hours by end of 2026.

Three commercialization signals ship in the same window: (i) Figure 02 General Brain secured a 50,000-unit BMW order in Q2 2026; (ii) Zhiyuan Longhe A2 lands a 10,000-unit agreement with BYD and SANY Heavy Industry; (iii) Physical Intelligence π0 licensing model delivered ARR of $150M by Q2 2026 (vs. $8M in 2025), with a $12B valuation — the closest analogue yet to “Android of robotics.” τ₀-VLA’s 40,115-hour pretraining plus TARS’s 1M+ hours plus Xiaomi’s 100,000 hours demonstrate the same data curve at different points on the timeline.

Why it matters: The EFM race is now the dominant frontier race in embodied AI, and the determining factor over the next 12 months is integrated real-world data + on-device inference economics rather than model architecture. Teams that can scale to 1M+ hours and run a >100B model on a Jetson Thor-class chip (2,000 TOPS INT8) will pull away from teams still collecting demos.

8. Tesla Concedes Optimus Has Three Unsolved Problems as Chinese Makers Ship ~90% of Global Humanoids

Category: Embodied Intelligence / Industry Reality Check Source: Business Insider (via ReadAboutAI), Omdia, WAIC 2026 coverage, Xiaomi event Date: 2026-07-28

On Tesla’s Q2 2026 earnings call, Elon Musk conceded Optimus still has three unsolved problems: reliable real-world intelligence, a functional dexterous hand, and scaled manufacturing without an established supply chain. The Fremont line that built Model S/X is being converted for Optimus production. Tesla’s $20K–$25K per-unit cost estimate depends on reaching million-unit production, with no confirmed timeline. Independent roboticists quoted by Business Insider called humanoid robots a “fantasy product,” noting self-driving took two decades to reach market after the core technology worked.

Omdia data shows Chinese manufacturers already dominated ~90% of humanoid shipments globally in 2025. WAIC 2026’s “deployment-state” theme demonstrated this lead operationally: Zhiyuan showed the A3 Ultra with a 700-TOPS embodied processor and high-DoF dexterous hand (selected as a “镇馆之宝”); Yushu demonstrated the DobotWAM mass-delivery model; AgiBot (Zhiyuan partner) released τ₀-VLA; Xiaomi unveiled Xiaomi-Robotics-1 with a published Scaling Law; Tencent open-sourced the Hy-Embodied VLA/VLM/RxBrain matrix; KEENON demonstrated humanoid + DINERBOT T9 hotel-laundry coordination. Forecasts for 2026 global humanoid shipments: 50,000+ units, +178% YoY, with Zhiyuan + Yushu together holding >50% commercial share.

Why it matters: Optimus’s admission is the cleanest articulation yet of how far ahead China’s hardware-plus-data-plus-supply-chain stack is for embodied AI. Western capital and frontier-model research are converging fastest on the software side of embodied intelligence, but the deployment side — dexterous hands, harmonic joints, large-scale real-trajectory data, robotics-as-a-service — has crossed decisively into Chinese industry hands. The relevant 2026–2027 question is no longer whether humanoid robots work; it is who can build a million-unit factory to make them.

Quick Takes

• τ₀-VLA published its Apache 2.0 model + code + paper the same day as the paper — pattern of “release-everything-day” continues from Xiaomi-Robotics-U0 and Kimi K3.
• GitHub removed Redis from its MCP Server pre-emptively before the spec even froze — first major infrastructure provider to publicly validate stateless MCP at scale.
• Anthropic Opus 5 is now the default model on Claude Max while Fable 5 stays as the high-cost frontier tier — pricing an entire Claude tier at $5/$25 per million tokens is the strongest sub-frontier pricing move of the year.
• Kimi K3 Decode benchmark on SGLang: ~113 tok/s single-card batch-1, ~423 tok/s with DSpark speculative decoding — first open model competitive with closed frontier models on interactive coding latency.
• α25-36% of European startup CEOs now say their biggest 2026 risk is AI agent evaluation cost, not model access — a perceptible shift from December 2025 surveys, driven by Codex Goal Mode and Claude Code subagent nesting.

Today’s Trend Lines

1. 2026-07-28 — The infrastructure underneath coding agents hit production maturity. MCP’s stateless revision, GitHub’s same-day support, Kimi K3’s open infra stack, GitHub Copilot’s Harness workflow — four independent moves on the same day. The interface layer is now the differentiator, the model itself is increasingly a commodity.

2. 2026-07-28 — Embodied AI’s center of gravity shifts to long-horizon reasoning and industrial deployment. τ₀-VLA’s test-time compute, TARS’s SAIL Award + million-hour data roadmap, Xiaomi-Robotics-1’s Scaling Law validation, and Yushu/Zhiyuan’s mass-delivery signals all point to a single inflection: the embodied field has matured past “single-step demos” into “long-horizon tasks on real production lines.”

3. 2026-07-28 — Open-weight frontier models reshape both economics and policy posture. Kimi K3’s free 2.8T weights + Anthropic’s clarification that it never advocated banning open weights + NVIDIA’s Open Secure AI Alliance together mark the point where “open vs closed” stopped being a community debate and became a national-security and policy question both Washington and Beijing are actively legislating around.

4. 2026-07-28 — China consolidates its embodied-AI lead on hardware + data + supply chain. Optimus’s three unsolved problems hit the same week as τ₀-VLA, Tencent’s full-stack, Xiaomi’s Scaling Law, Zhiyuan’s A3 Ultra, and Yushu’s mass delivery. Western capital still leads on foundation-model research; Chinese industry now leads on what to do with the resulting intelligence.

Compiled by EAIDaily automation. Sources: AI HOT (aihot.virxact.com), Anthropic Newsroom, GitHub Blog, Moonshot AI X / blog, LMSYS Blog, Shanghai Innovation Institute, TARS PRNewswire, Tencent Cloud, WAIC 2026 coverage, The Register, Business Insider, 财联社, Wall Street CN, 量子位, 机器之心, TechStartups, ReadAboutAI. View historical archives at EAIDaily_YYYY-MM-DD.md.