DeepSeek releases native multimodal vision model, researchers claim hackers use it for attacks, and Pinduoduo Q2 adjusted profit beats expectations.
— V4 Roadmap Fills Key Gap in Image-Text Understanding According to TMTPost, DeepSeek launched a Vision model supporting native multimodality, part of the subsequent V4 version capabilities. Founder Liang Wenfeng previously stated in investor exchanges that multimodality is a component rather than the main line for intelligence ceilings, but subsequent V4 versions would natively support it. Now this judgment has materialized. Native multimodality means image and video understanding capabilities enter the model base directly, no longer adapted via external plugins, allowing C-end applications and API product matrices to call them synchronously. For the open-source ecosystem, DeepSeek's multimodal weights have long been one of the community's most important bases. This update will drive a round of updates in downstream applications and fine-tuning ecosystems. Domestic large model competition thus extends from text to multimodality, heating up the battle for C-end assistant entry points. Multimodal inference consumes more compute, making marginal changes in inference-side hardware demand worth tracking. For the industry chain, multimodal weights require significantly higher VRAM and inference throughput than pure text, further increasing the unit value of compute demand. Adaptation progress of domestic inference chips and accelerator cards will also receive more attention. From a product rhythm perspective, the release of the Vision model means DeepSeek's support for C-end multimodal applications is ready. Image understanding and visual Q&A products can directly call native capabilities without stitching external vision modules. A batch of lightweight multimodal applications is expected to land quickly.
— AI Security Agenda Extends from Content Side to Offense-Defense Confrontation Bloomberg cited researchers' reports stating that hacker groups with Chinese backgrounds have begun using open-source large models like DeepSeek to improve the efficiency and automation levels of cyberattacks. Signs of model assistance appeared in attack writing, obfuscation, and targeting stages. The spillover of open-source model capabilities into attack scenarios expands AI security discussions from content safety to offense-defense confrontation. National regulations on acquiring and using open-source weight models may tighten. For the security industry, demand for red team detection, threat intelligence, and endpoint protection rises accordingly. For the open-source ecosystem, controversies over model capability misuse may affect DeepSeek's overseas ecosystem expansion pace. This will be a long-term balance issue between open source and security. From an investment perspective, AI security shifts from compliance cost to real demand. Order visibility in sub-sectors like threat intelligence, model red teaming, and identity security is improving. Meanwhile, discussions around export and hosting policies for open-source models will increase. Classification management of model weights by countries may become the next regulatory agenda item, raising compliance costs for the open-source community.
— Inventory Concerns Emerge Under AI Boom Bloomberg reported that Eoptolink (Innolight), one of the biggest beneficiaries of AI optical module demand, stockpiled heavily during the boom cycle, causing inventory scale to climb sharply. The optical module segment has maintained high prosperity for multiple consecutive quarters. The market begins to reassess inventory digestion rhythm and price hike sustainability. Inventory surge has two interpretations: locking capacity for delivery preparation, or early signals of demand overdraft. Bulls and bears can find arguments in both. A-share optical module duo stocks adjust at highs. Sector sentiment shifts from unilateral optimism to divergence. If subsequent inventory destocking falls short of expectations, valuation center has room for downward revision. If demand continues to exceed expectations, stockpiling instead constitutes delivery advantage. Divergence itself shows compute hardware trading has entered the verification period of prosperity realization. On the capital side, sector valuations are at highs after continuous rises. Any inventory and price signals will be amplified. Holding volatility is unlikely to drop in the short term. From position structure, optical modules are one of the heaviest allocations for institutions in AI hardware. Reflexivity of crowded trading means stock price sensitivity to negative information rises significantly. Marginal changes in single-product inventory data could trigger rebalancing.
— AI Server Export Compliance Risks Heat Up Ars Technica reported that a senior Nvidia manager is allegedly linked to Supermicro's suspected smuggling plan to funnel AI servers to China, bypassing US export controls. US export controls on AI chips to China enter deep implementation waters. Executive involvement reinforces compliance risk narratives and indirectly confirms the huge demand gap for advanced compute in the Chinese market. For Nvidia, regulatory and reputation uncertainty rises, putting short-term stock pressure. For domestic compute industry chain, substitution logic gains new support. Verification and procurement rhythm of domestic GPUs may accelerate due to this. For Supermicro, the event adds to previous governance issues. Repairing compliance costs and customer trust for its server business still takes time. For other suppliers in the export control chain, this case is also a warning. Manufacturers with unsound compliance systems face risks of exclusion from mainstream client lists.
— Another Highlight in Chinese Concept Stock Earnings Season Pinduoduo released its Q2 2026 earnings. Adjusted profit exceeded market expectations. Quarterly revenue 112.4 billion yuan, profit 27.2 billion yuan, benefiting from the release of effects from merchant support and platform governance upgrades. Under dual pressure of Temu overseas expansion and domestic e-commerce competition, profitability resilience
📰 EastVoice — today's AI headlines, decoded
Five minutes a day, the global AI shift →