物界前沿【East Voice】2026-08-01 China AI Daily Digest
East Voice · China AI Daily Digest
August 1, 2026 · Saturday
Editor’s Note: DeepSeek V4-Flash rewrote the pricing playbook with post-training-only gains, while China’s humanoid robot sector hit 40,000 units in H1. This is the weekend edition — a curated selection of QC-verified stories.
1. AI Foundation Models (LLM)
1. DeepSeek V4-Flash Goes Live: Agent Capabilities Surge 7.5x, MIT-Open-Sourced at Rock-Bottom Pricing
Date: 2026-07-31
DeepSeek launched the official public beta of V4-Flash-0731 API. Same 284B total parameters / 13B active / 1M context window — but post-training alone drove DeepSWE from 7.3 to 54.4 (+7.5x) and Terminal Bench to 82.7, rivaling Claude Opus 4.8. MIT-licensed weights. Priced at just ¥1/M input tokens and ¥2/M output tokens, with cache hits at ¥0.02. Native Responses API support with Codex integration. Concurrently announced 1GW compute expansion in Ulanqab.
Source: DeepSeek Official Changelog, Kingy AI, NetEase
Editor’s Take: Architecture unchanged, post-training only, yet Agent capabilities crush their own Pro preview — DeepSeek just proved “post-training is where the magic happens.” When Flash outperforms Pro, the “you get what you pay for” pricing logic is permanently disrupted.
2. Tencent Hunyuan Open-Sources AngelSpec Speculative Decoding Framework — Up to 2.4x Inference Speedup
Date: 2026-07-30
Tencent Hunyuan open-sourced speculative decoding framework AngelSpec with accompanying weights and code, supporting MTP and block-parallel draft model training across six architectures. New drafter DFly achieves 1.98–2.40x average speedup, peaking at 2.86x.
Source: Toutiao
Editor’s Take: Inference acceleration is one of the core levers for reducing LLM deployment costs — Tencent continues to invest heavily in open-source infrastructure.
3. Tencent Hunyuan’s Hyra Solves 50-Year Math Problem, AI Research Agent Enters Global First Tier
Date: 2026-07-31
Tencent Hunyuan’s research agent Hyra (based on Hy3 model) solved an open problem in additive combinatorics that had been unsolved for over half a century, running for approximately 24 hours. It improved the best known result from ~1.14 to 1.21 and provided a Lean 4 formal proof.
Source: Internet Thought, NetEase
Editor’s Take: From “assisting human research” to “independently solving math problems” — Chinese AI has entered the global first tier in “using AI to do science.”
2. AI Software (Application / SaaS)
1. ByteDance Releases Trae 3.0: From “Code Completion” to “Full-Process Agent”
Date: 2026-07-31
ByteDance released Trae IDE 3.0 on WAIC 2026 closing day, positioned as “full-process agent” rather than “code completion tool.” Upgrades include Solo Mode 2.0, multi-agent collaboration, cross-file refactoring, terminal agent, and model switching (GPT-5.6, Claude Opus 5, etc.). 5M MAU, 70% coding efficiency improvement.
Source: Toutiao
Editor’s Take: Benchmarking GitHub Copilot (~8M MAU) and Cursor (~3M), Trae reached 5M MAU in under a year. Chinese tools are collectively leaping from “code completion” to “full-process agent” — and leaping faster than Silicon Valley.
2. Meitu Paid Users Exceed 18.44M, AI Productivity App ARR Up 47.8%
Date: 2026-07-31
Meitu (1357.HK) issued H1 earnings guidance: paid subscribers exceed 18.44M (record high), imaging & design revenue up 30.9% YoY to ~¥1.8B, AI productivity app ARR up 47.8% YoY. Adjusted net profit up 36%–40% YoY.
Source: Toutiao, ChinaNet
Editor’s Take: AI productivity tools are creating new growth curves for Chinese SaaS companies — Meitu’s pivot from “beauty filter” to “AI design platform” is now data-validated.
3. MiniMax H3 Open-Source Omnimodal Video Model: 2K Native Stereo at 1/3 Competitor Cost
Date: 2026-07-31
MiniMax launched H3 multimodal generation model — unified understanding of text, image, video, and audio, generating up to 2K resolution, 15-second video with native stereo audio. Priced at less than 1/3 of mainstream models at 2K. Weights to open-source on ModelScope on August 3.
Source: Internet Thought, NetEase
Editor’s Take: The video generation track enters the “open-source democratization + extreme cost-performance” phase — MiniMax’s price-disruptor strategy is reshaping market dynamics.
4. Amap (Gaode) Launches ABot Embodied System Full-Stack Upgrade with Five Core Models
Date: 2026-07-31
Alibaba’s Amap announced a full-stack upgrade of its ABot embodied system, releasing five core models including ABot-N1. Integrating perception, decision-making, execution, and memory — the world’s first full-stack system fusing world model, foundation model, and embodied agent.
Source: AIbase
Editor’s Take: From navigation to robotics — Amap’s ABot marks the “world model + embodied intelligence” fusion route landing in China.
3. Humanoid Robots
1. MIIT: China Humanoid Robot Production Exceeds 40,000 Units in H1, Full-Year Expected to Surpass 100,000
Date: 2026-07-30
MIIT Vice Minister Ke Jixin disclosed that China’s humanoid robot production exceeded 40,000 units in H1 2026, with full-year output expected to surpass 100,000. BYD’s humanoid robot to debut in August; XPeng IRON enters small-batch trial production. H1 saw 288 funding rounds totaling ¥46B+ in robotics.
Source: Daily Economic News
Editor’s Take: From “dancing on stage” to “shipping in batches” — official MIIT data confirms the humanoid robot industry has reached a scale-up inflection point.
2. Beijing Yizhuang ESCC Conference Launches “Gene Factory” — Embodied Supply Chain 2-Hour Response
Date: 2026-07-30
The 2026 ESCC Embodied Intelligence Supply Chain Conference convened in Beijing Yizhuang with 400+ companies. The “Gene Factory” rapid-response supply chain platform debuted — as fast as 2-hour response for components/pilot/OEM needs. Yizhuang has built China’s first humanoid robot pilot verification platform and first 10,000-unit-scale factory.
Source: Toutiao
Editor’s Take: The key to embodied intelligence scale-up isn’t the model — it’s the supply chain. “2-hour response” signals the robotics industry chain is replicating the smartphone industry’s maturity playbook.
3. Beijing Rolls Out 3-Year Humanoid Robot Action Plan (2026–2028), Prioritizing “4D” Scenarios
Date: 2026-07-30
Beijing will issue a three-year action plan for humanoid robot industry development, prioritizing Dull, Difficult, Dangerous, and Dirty (“4D”) scenarios. Guangdong Shunde established China’s first embodied intelligence bureau. MIIT and SASAC jointly launched the 2026 Real-Scene Training Special Action.
Source: Beijing Youth Daily
Editor’s Take: The “4D” scenario-first strategy is pragmatic — replace dangerous, repetitive, dirty work first, solving real pain points while lowering social acceptance barriers.
4. Agibot WITA-Omni Tops DailyOmni Benchmark — Embodied Multimodal Outperforms General LLMs
Date: 2026-07-30
Agibot’s WITA-Omni Preview scored #1 on the global embodied multimodal understanding benchmark DailyOmni (85.21 points), topping six sub-categories including audio-video alignment, event sequencing, and long-video understanding — surpassing Qwen, Gemini, Doubao, and NVIDIA models.
Source: Toutiao
Editor’s Take: A robot hardware company outperforming general LLMs on multimodal benchmarks — this bridges the last gap for face-to-face robot interaction, pointing toward commercial deployment.
5. Suzhou AI Product Expo Opens with Nearly 400 Exhibitors Showcasing Embodied Intelligence
Date: 2026-07-30
The 2026 AI Product Application Expo opened in Suzhou with nearly 400 exhibitors showcasing domestic LLMs and full-scenario embodied robots. The concurrent World Intelligent Industry Expo featured 80+ companies and 150+ products. Galileo robot dogs secured ¥10M+ in intent orders.
Source: Toutiao
Editor’s Take: Two major expos running simultaneously with record exhibitor numbers — industry consensus has shifted from “single-scenario pilots” to “cross-scenario commercial scale-up.”
6. Embodied Data Infrastructure Becomes New Funding Hotspot: Delta Intelligence Raises ¥500M, Guanglun Becomes Unicorn
Date: 2026-07-31
Delta Intelligence closed ~¥500M Angel++ round, launching Delta D1 head-mounted full-body data capture device. Guanglun Intelligence surpassed ¥15B valuation, becoming the world’s first embodied data unicorn. Industry calls 2026 the “year of embodied intelligence data scale-up.”
Source: Toutiao
Editor’s Take: The core bottleneck for embodied intelligence isn’t models — it’s data. The industry is shifting from “competing on models” to “competing on data infrastructure.”
7. Humanoid Robots Flood into Auto Manufacturing — Densest Deployment Scenario
Date: 2026-07-30
Figure enters BMW production lines; UBTECH enters NIO and BYD; Agibot enters SAIC-GM battery lines; Qianxun “Xiao Mo” scales at CATL. Junpu G2 robot completed 8-hour, 2,283-task zero-error test at a leading Chinese 3C manufacturer.
Source: Toutiao
Editor’s Take: Auto manufacturing is becoming the densest deployment scenario for embodied intelligence — flexible upgrade demand and labor shortages together create rigid demand.
8. Unitree Robotics Pursues STAR Market IPO, Catalyzing Humanoid Robot Supply Chain
Date: 2026-07-31
Unitree Robotics is pursuing a listing on China’s STAR Market (科创板). Industry observers believe a leading company’s A-share listing will catalyze upstream development in reducers, servo motors, sensors, and the broader embodied intelligence supply chain.
Source: Toutiao
Editor’s Take: The STAR Market may soon welcome the “first humanoid robot stock” — capital market recognition will further accelerate industry chain investment.
东声 East Voice — Physix Frontier’s China AI column for global readers
东声西达，中国视角。
© 2026 Physix Frontier · All rights reserved