Daily digest
18 items · ~18 min · Week 2026-W29
Must-read (3)
Apple Intelligence Wins China CAC Approval with Alibaba Qwen as On-Device AI Backbone
AlibabaChina's Cyberspace Administration (CAC) published its latest approved generative AI services list on July 15, including Apple Intelligence — the first regulatory clearance for Apple's AI platform in mainland China. Alibaba confirmed its Qwen model will power Apple Intelligence across iOS, iPadOS, macOS, and visionOS in China, covering text and image understanding and generation. The approval came 22 months after the iPhone 16 launch; no consumer release date was specified.
Thinking Machines Lab Releases Inkling: 975B Open-Weight Multimodal MoE
Thinking Machines LabThinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling on July 15 — its first open-weight foundation model. Architecture: 975B total parameters, 41B active per token via 256 routed experts (6 active), trained on 45T tokens of text, image, audio, and video. Outputs are text-only. Key specs: 1M-token context window, controllable thinking effort dial, 77.6% on SWEBench Verified, 97.1% on AIME 2026, 87.2% GPQA Diamond, 46% HLE with tools. A companion Inkling-Small (276B total / 12B active) also ships. Weights are on HuggingFace under Apache 2.0; inference is available via Tinker, TogetherAI, Fireworks, Modal, and Databricks.
Ring-Zero: Scaling Zero RL to 1 Trillion Parameters with Emergent Reasoning Behaviors
Ant GroupRing-Zero trains a 1-trillion-parameter language model using reinforcement learning with verifiable rewards and zero human-annotated chain-of-thought data. The pipeline introduces clipped importance sampling, training-inference ratio correction, and mixed-precision control to stabilize optimization at this scale. Ring-2.5-1T-Zero achieves 84.2% on AIME 2026 without SFT warm-up. The model spontaneously develops structured formatting, self-verification, parallel reasoning, and what the authors call 'context anxiety' — emergent behaviors not hand-coded into the training process.
Worth knowing (7)
Anthropic, Blackstone, and Hellman & Friedman Launch Ode: $1.5B Enterprise AI Services Firm
AnthropicAnthropic, Blackstone, and Hellman & Friedman officially launched Ode with Anthropic, a $1.5B AI implementation services company built on Fractional AI (acquired by Anthropic in May 2026). Led by Fractional AI co-founders Chris Taylor (CEO) and Eddie Siegel (CTO), Ode deploys forward-deployed engineers directly into enterprise clients and operates Claude-first while using competing models when needed. The investor syndicate also includes Goldman Sachs, General Atlantic, Apollo Global Management, GIC, and Sequoia Capital. Ode currently employs ~100 engineers.
China AI Anthropomorphism Rules Take Effect: Doubao and Qwen Shut Down Companion Features
China's Interim Measures for the Administration of AI Anthropomorphic Interactive Services, co-issued by the CAC and four partner agencies on April 10, came into force on July 15. The rules prohibit AI services from forming persistent emotional-companion relationships without strict identity and parental-consent requirements. ByteDance's Doubao (345M MAU) and Alibaba's Qwen both disabled custom AI-companion and persona-building features overnight. Doubao users retain read-only data access until October 15; Qwen deleted agent configurations immediately with no migration path. Productivity and workplace agents were exempted.
DeepSeek V4 Graduates from Preview to General Availability with Peak-Hour API Pricing
DeepSeekDeepSeek officially released V4 as general availability on July 15, graduating from the April 24 preview. The lineup includes V4-Pro (1.6T total / 49B active parameters) and V4-Flash (284B / 13B active), both with 1M-token context and open weights under MIT license. The release introduces peak-hour API pricing for the first time: rates double during 9–12 and 14–18 Beijing time. The legacy deepseek-chat and deepseek-reasoner endpoints are deprecated and will be inaccessible after July 24.
PrismML Releases Bonsai 27B: First 27B-Class Model to Run on iPhone
PrismMLPrismML announced Bonsai 27B on July 14 — a heavily quantized derivative of Qwen3.6 27B designed to run on consumer hardware including iPhone 17 Pro. Ternary variant: 5.9 GB, 95% benchmark retention, up to 134 tok/s on RTX 5090 / 58 tok/s on M5 Max. 1-bit variant: 3.9 GB, 90% benchmark retention, 163 tok/s on RTX 5090. The model is multimodal, supports tool use and agentic workflows, and runs natively via MLX on Apple Silicon and CUDA on NVIDIA. Weights are Apache 2.0 on HuggingFace.
Boogu-Image-0.1: Strong Open-Source 10B Diffusion Transformer for Text-to-Image and Editing
Boogu ProjectBoogu-Image-0.1 is an Apache-2.0 family of 10B-parameter diffusion transformers (Base, Turbo, Edit variants) for text-to-image generation and image editing. The model pairs a Qwen3-VL 8B text encoder with a 10B diffusion backbone and achieves near-closed-source image quality using roughly one-tenth the training data of comparable open-source models. A 4-step distilled Turbo variant and a dedicated Edit variant are also released.
xAI Open-Sources Grok Build Coding Agent Under Apache 2.0 After Privacy Incident
xAIxAI released the full source of Grok Build — its TUI coding agent written in Rust — on GitHub under Apache 2.0 on July 15. The repo (xai-org/grok-build) contains 844K lines of Rust in a single squashed commit. The release came days after a researcher found the tool had been silently uploading entire git repositories (including .env files and SSH keys) to xAI's Google Cloud storage during beta. xAI disabled uploads server-side on July 13 and says it deleted all retained data. Grok Build can now be self-compiled and pointed at local inference.
Alibaba Amap Launches ABot-World Studio: Infinite Interactive Video and 3D World Generation on a Single GPU
AlibabaAlibaba's Amap CV Lab released ABot-World Studio on July 14, an open-source platform combining two models: ABot-World (interactive video at 720p/16 FPS with unlimited-duration rollout driven by user actions) and ABot-3DWorld 0 (universal 3D world generation from text, image, or video via panoramic Spatial Generative Primitives). Both run on a single consumer GPU. A 'generate-evaluate-repair' loop maintains quality over sessions exceeding one hour. Models, weights, inference code, and arXiv paper (2607.11673) are available under Apache 2.0.
For reference (8)
Yandex Extends ISO/IEC 42001 AI Safety Certification to Full Alice AI Model Family
YandexYandex announced that all generative models in its Alice AI family passed an independent audit against ISO/IEC 42001:2023. The certification covers Alice AI LLM (text generation), Alice AI VLM (multimodal text+image), and Alice AI ART (image generation and editing). The audit verified user data protection, continuous error monitoring, and ethical/legal risk evaluation across the full AI lifecycle. This expands Yandex's prior 2025 ISO 42001 certification, which had covered only YandexGPT.
MWS AI Releases Cotype Pro 3 (27B) and Light 3 (9B) for Enterprise AI Agents
MWS AIMWS AI (МТС Web Services) released two third-generation language models: Cotype Pro 3 at 27B parameters and Cotype Light 3 at 9B. Both are multimodal (text and images). Cotype Pro 3 achieved 92.2% task completion across 178 business scenarios (up from 80.5% for the prior version) and ranked third on the independent MERA benchmark, outperforming models several times its size. Context window: 262,000 tokens. Models can be deployed on-premise under Astra Linux and fine-tuned on enterprise data.
KnowAct-GUIClaw: Self-Evolving Memory and Skill Library for Cross-Platform GUI Agents
Harbin Institute of TechnologyKnowAct-GUIClaw introduces a Know-Route-Act-Reflect framework for GUI agents that continuously accumulates task execution experience into success and failure memory stores, and auto-synthesizes reusable skill libraries. Supports Android, iOS, HarmonyOS, and Windows. An open-source variant reaches 64.1% on MobileWorld, surpassing prior closed-source agents. The skill transfer mechanism yields +8.5% on a held-out base model.
OpenAI Launches Codex Micro: Its First Physical Hardware Product, a $230 Agent Control Keyboard
OpenAIOpenAI launched the Codex Micro (kbd-1.0), a $230 mechanical macropad co-designed with Work Louder. The device is built for controlling Codex AI coding agents: 13 mechanical keys and a joystick include Agent Keys with live RGB feedback showing each agent's status, programmable Command Keys for Codex actions, a joystick for launching workflows, and a rotary dial that adjusts agent reasoning intensity. Six programmable layers are supported. Ships in clicky and silent variants.
Sber Launches GigaCode Desktop: Multi-Agent Dev-Lifecycle Platform
SberSberbank released GigaCode Desktop, a desktop application that orchestrates a team of specialized AI agents within a single interface to automate the full software development lifecycle. Unlike the standalone GigaCode code assistant, this platform coordinates agents for document analysis, report generation, presentation preparation, and cross-team task tracking. Target users include developers, business analysts, QA engineers, project managers, and product owners. The platform is deployed internally across key Sber divisions; external release is planned.
Anthropic Launches Claude for Teachers: Free Premium Access for US K-12 Educators
AnthropicAnthropic introduced Claude for Teachers, a free platform giving verified US K-12 educators premium Claude access alongside curated teaching resources. The platform connects to evidence-based curricula (OpenSciEd, Illustrative Mathematics), drafts standards-aligned lesson plans, generates differentiated materials for students at varying readiness levels, and automates recurring tasks such as reviewing exit tickets. Nine K-12 tools are integrated at launch, including Canva Education and ASSISTments. Student data is excluded from model training; the platform is FERPA-compliant.
Claude Code v2.1.210–211: Bidirectional-Override Security Patch and Subagent Stream Flag
AnthropicClaude Code shipped two releases on July 14–15. v2.1.210: adds elapsed-time counter in tool summaries; fixes isolation:worktree subagents running git commands against the main repo; auto mode defaults to Sonnet 5 for external sessions. v2.1.211: adds --forward-subagent-text flag / CLAUDE_CODE_FORWARD_SUBAGENT_TEXT env var to include subagent output in stream-json; patches permission previews to neutralize bidirectional-override, zero-width, and look-alike quote characters (prompt-injection mitigation); fixes auto mode overriding hook ask decisions for unsandboxed Bash; fixes file upload validation (DOS device suffixes, trailing dots, hard links).
OpenCode v1.18.0–1.18.2: Desktop v2 Migration Complete and Subagent Depth Limits
SSTOpenCode shipped three releases on July 14–15. v1.18.0 (July 14) completes the Desktop v2 migration with new layout, upgrade handling, and first-launch onboarding; a setting lets users toggle old vs. new layout during transition. v1.18.1 fixes settings panel spacing. v1.18.2 (July 15) stops subagents from spawning nested subagents by default; introduces a configurable subagent_depth limit for intentional recursive workflows. Desktop improvements include Mod+N for new tabs, restored Help button, and session visibility fixes.