Tencent open-sourced Hy4 preview, a 770B MoE that beat GLM-5.3 and Kimi K3 in its own blind test
Tencent released and open-sourced Hy4 preview on August 28, a mixture-of-experts model with 770 billion total parameters, 49 billion active, and a context window past one million tokens. The release numbers matter more than the architecture: in Tencent's own blind evaluation, 163 internal experts scored it 2.99/4.00 across 203 engineering tasks, edging out GLM-5.3 (2.92) and Kimi K3 (2.94). It ships inside WorkBuddy and CodeBuddy (both CN and international builds), Yuanbao and ima, with a two-week free window on the first two, and API pricing is set at $0.834 per million input tokens and $2.501 per million output. The license is Apache 2.0.
The part that reads like a research note rather than a product launch is the self-improvement claim. Tencent says Hy4 preview participated in optimizing its own training methods, data strategy, evaluation frameworks and low-level operators, then autonomously analyzed its own inference bottlenecks and raised end-to-end throughput by 31.8%. That is a recursive self-improvement loop stated plainly, and it lines up with Anthropic's automated alignment researchers from last week. My read: the blind-test margin over GLM-5.3 and Kimi K3 is small, and Tencent is upfront that this is an early version with room to grow on both pre- and post-training. The cadence is the signal — Hunyuan has shipped a major version roughly every two months since rebuilding its infrastructure in February, and the next Hy4 batch is already scheduled.
— Tencent (official) · Tencent Cloud · People's Daily
🔗 Tencent: Tencent releases and open-sources Hy4 preview · Tencent Cloud on the 770B MoE details · People's Daily on the launch
Anthropic's IPO is in its final stretch: a prospectus after Labor Day and a $130 billion target raise
Anthropic plans to file its prospectus after the US Labor Day holiday (September 7) and list in late September or early October, according to 财联社 (Cailianshe) and follow-on reports, with a target raise of at least $130 billion. That would more than double the record SpaceX set in June ($86 billion) and make this the largest IPO in history. The valuation story is the one that pushed the company to a $965 billion post-money in May's $65 billion Series H: an annualized revenue run rate of $65 billion at the end of July, Q2 revenue above $11.5 billion (roughly 14x the year-ago quarter), and the first quarter of positive adjusted operating profit. Underwriters are reported as Goldman Sachs, JPMorgan and Morgan Stanley, and a revolving credit facility of more than $10 billion is being arranged.
Two structural details separate this from SpaceX and Cerebras. Anthropic is considering letting existing shareholders sell into the IPO — a secondary component alongside new shares — and is weighing lockups of more than 180 days for at least some holders to limit post-listing selling pressure. It also cleared a legal obstacle this week: a federal judge ruled the Pentagon's blacklisting of Anthropic unlawful, removing a risk from the offering. My read: the prospectus will answer the question that matters more than the raise size — how much of the $65 billion run rate is durable enterprise revenue versus compute-credit arithmetic. That document, not the roadshow number, is what allocators should wait for.
— 财联社/Cailianshe · 新浪财经 · Nasdaq
🔗 Cailianshe on the Labor Day prospectus timing · Sina Finance on the $130B raise and shareholder sales · Nasdaq on the S-1 and Amazon's stake
CXMT shipped the world's first commercial LPDDR6, and it went into a Xiaomi foldable
China's CXMT (长鑫科技) announced on August 29 that its self-developed LPDDR6 memory has entered mass production, with the first commercial deployment on Xiaomi's 18 Fold foldable flagship. That makes CXMT the first company to ship LPDDR6 in a product, breaking a launch sequence that Samsung, SK Hynix and Micron have controlled for every previous memory generation. The chip runs at a peak 12,800 Mbps with up to 16GB per chip, and it pairs with Xiaomi's Xuanjie O3 SoC, the first mobile processor designed for LPDDR6. The JEDEC standard itself only came out in July 2025, so CXMT went from standard to shipping silicon in about a year.
LPDDR6 is being framed as the memory that unlocks on-device AI — more bandwidth for local model inference, extending beyond phones into PCs, smart cockpits and AI data centers. The business context is as loud as the technology. CXMT reported first-half net profit of 77.6 billion yuan, turning around from a loss, and its roughly 4 trillion yuan market cap makes it the most valuable stock on the A-share market. Xiaomi's Lei Jun publicly congratulated CXMT and confirmed the co-design. My read: the significance is less the benchmark numbers and more the sequencing — Chinese memory, Chinese SoC and a Chinese flagship phone brought a new memory standard to consumers before the incumbents did. One commercial launch does not overturn a decade of DRAM dominance, but the order of events is new.
— CXMT (official) · 证券时报 · 科创板日报
🔗 证券时报 on the LPDDR6 mass production · 科创板日报 on the specs and the Xiaomi tie-up · 北京商报 on the world-first claim
A dual-arm robot now makes DQ Blizzards in Shanghai — 55 steps, tactile sensing, no store remodel
On August 29 a white dual-arm robot started a shift at a Dairy Queen on Wujiang Road in Shanghai, making Blizzard ice cream end to end: pulling a paper cup from a sanitizing tray, aligning the cup ring, adding toppings, stirring the thick mix without spilling, and turning the cup upside down to prove the "no-spill flip." That is the full 55-step workflow, run without human intervention, in a store where nothing was modified — the same equipment, ingredients, supply chain and operating standards as any other DQ. The robot takes about 6.5 minutes per cup, roughly half the speed of an experienced human, and the store runs three robots on 12-hour shifts, year-round. The company is Sharpa, founded in late 2024 by the three co-founders of LiDAR maker Hesai, which just disclosed cumulative funding of more than 4.5 billion yuan ($630M+) at a post-money valuation above 22 billion yuan, with Alibaba, Meituan, Tencent, JD.com, Transsion and Sequoia China among the backers.
The detail that separates this from earlier robot-barista demos is the tactile layer. Sharpa says 98% of the 55 steps depend on tactile sensing — reading friction when pulling the cup, adjusting grip force in real time during high-speed mixing — through its Sharpa Wave hand, which has 22 active degrees of freedom and more than 1,000 tactile sensing units, paired with the CraftNet foundation model. It also deliberately did not train Blizzard-making as one closed procedure; the task is decomposed into reusable skills (opening cabinets, retrieving, aligning, scooping, mixing, pouring, handing over) so the same model can move to other venues. Co-founder Li Yifan is candid that the first-generation robot is not yet profitable and that the industry's early deployments are unlikely to show positive ROI in the short term. My read: the honest benchmark is the 6.5 minutes and the 50% efficiency — this is a deployment built to collect data and prove reliability, not to make money yet, and that is the right order of operations for dexterous manipulation.
— 新华财经 · 每日经济新闻 · The Insight Asia
🔗 新华财经 on the DQ robot restaurant opening · 每日经济新闻 on the economics of a robot employee · The Insight Asia on the funding round
Europe's first humanoid robot plant opened in Serbia — a Chinese auto-parts maker and AgiBot
Europe's first mass-production humanoid robot plant opened on August 29 in Šabac, Serbia, a partnership between Chinese auto-parts maker Minth Group and Shanghai's AgiBot Innovation (智元机器人). Serbian President Aleksandar Vučić attended and greeted the first robot off the line, which wore traditional Serbian dress and danced the kolo. The first phase is a €20 million investment at Minth's existing Majur facility, targeting more than 5,000 robots a year, and the next phase is a Robotics Industrial Park in Inđija worth around €200 million with planned capacity of up to 20,000 humanoid robots and robot dogs annually for European and global markets. About 200 people will work in the new division initially, and 80 robots from the plant are slated to greet visitors at Expo 2027 in Belgrade.
The partner choice is the strategic tell. AgiBot delivered roughly 8,400 humanoid robots in the first half of 2026, about 44% of global shipments, ahead of Unitree's 31%; global H1 shipments were around 19,100 units, up 272% year over year, with China accounting for 97% of production. Minth, a Hong Kong-listed auto-components group with 27,400 employees and factories in 15 countries, already employs about 2,200 people in Šabac as the city's largest employer. The Serbian Development Agency (RAS) frames the plant as combining Serbian industrial infrastructure with Chinese physical-AI technology, and Vučić says the robots will carry a "Made in Serbia" label. My read: this is supply-chain and political capital more than technology transfer — Chinese embodied AI is embedding itself next to the EU industrial map, and the 20,000-unit target in Inđija is the number to watch, not the opening-ceremony demo. Whether European buyers accept Chinese-owned humanoid capacity is the real test.
— RAS (official) · Srpske Novine · Open4Business
🔗 Serbian Development Agency: Minth launches Europe's first mass production of humanoid robots · Srpske Novine on the opening and the 20,000-unit plan · Open4Business on the €20M phase one
OpenAI open-sourced Harness, the engine behind Codex — six times fewer tokens
OpenAI open-sourced Harness, the engine that runs Codex, on August 20 under Apache-2.0. The release is three layers: codex exec, the CLI for one-off scripted tasks; the Codex SDK for embedding the agent into applications; and app-server, the core execution server. Harness is the part that turns a model into an agent — task decomposition, long-conversation memory, real-time event streaming, tool invocation, interruptibility and human-in-the-loop approval — and OpenAI is pitching it as infrastructure any company can build agents on, in direct competition with Anthropic's Claude Agent SDK.
The numbers attached to the release are the reason developers should care. OpenAI says optimizing the harness alone raised GPT-5.6 Sol's ARC-AGI-3 score from 13.3% to 38.3% and cut token consumption to one-sixth for comparable tasks. A tax-preparation pilot processed 7,000 returns and cut prep time by about a third, and Cisco is already using the Codex SDK inside its Cloud Control platform. My read: the interesting signal is that OpenAI is now shipping the orchestration layer as open infrastructure rather than keeping it proprietary — the same pattern Anthropic and DeepSeek have been pushing from the other side. The benchmark gains say harness design is a bigger lever than most people give it credit for, and the token economics say the cost floor for agentic coding keeps dropping, which is pressure on every closed alternative.
— OpenAI (official/GitHub) · Open Source For You · Cynoteck
🔗 OpenAI Codex on GitHub · Open Source For You on the release and the numbers · Cynoteck on the 6x token cut and the platform play
Mistral's Agentic Search turns retrieval into a loop, and FinanceBench jumps from 26.7% to 86%
Mistral released Agentic Search on August 20, a retrieval layer that replaces one-shot RAG with a multi-step loop. Instead of pulling a fixed set of chunks and answering in a single pass, the model gets five tools that behave like a file system — search, open, navigate, read and grep — and can refine a query, open a specific document, jump to a table, read it and check the claim against another source before answering. The gains are large exactly where one-shot retrieval fails: on FinanceBench (150 questions over 368 SEC filings, about 147 pages each), correctness rose from 26.7% to 86%; on OfficeQA Pro (696 scanned Treasury Bulletins), from 6.3% to 51.9%. Latency and token use also fell — p90 dropped from 255 to 154 seconds, tokens by up to a third — because the loop stops fetching material it does not need.
The deployment detail is what makes it a European story. The Search Toolkit ships as open modules (ingestion, embedding, indexing) that run on the customer's own hardware behind the firewall, which is the answer to the data-residency clause that keeps European banks, hospitals and ministries off cloud AI. It also fits Mistral's physical build-out: a 10MW inference facility at Les Ulis near Paris is due this quarter. Two caveats belong on the record: the benchmark figures are Mistral's own and have not been independently reproduced, and both test sets are financial-document-heavy, the setting agentic retrieval flatters most. My read: the accuracy jump matters, but the strategic point is the business model — American frontier labs sell retrieval as a managed service where the margin lives, and Mistral is giving away the plumbing and selling the model and the hosting. In any tender with a residency clause, that is a structural advantage.
— Mistral (official) · AI in Europe · AI Reiter
🔗 Mistral: Introducing Agentic Search · AI in Europe on the on-premises angle · AI Reiter on what the numbers actually show
Top comments (0)