DeepSeek's Secret Self-Developed Inference Chip
Related
News: Zhipu is evaluating its self-developed custom AI chip
According to a report by The Information, multiple sources revealed that Zhipu, a leading Chinese AI company, is considering designing its own AI chip. It is understood that with the rapidly increasing demand for its GLM series of large-scale models, coupled with increasing constraints on computing resources, Zhipu has recently made initial contact with some domestic chip design companies to explore the possibility of jointly developing a customized AI processor. The report points out that for large-scale AI model developers, developing their own chips aims to achieve hardware and software synergy optimization, improve computing efficiency, and reduce dependence on external GPU suppliers. Besides Zhipu, another domestic large-scale model company, DeepSeek, has also recently been reported to be simultaneously advancing its self-developed chip project focused on inference scenarios. Globally, top AI developers such as OpenAI (which has announced its self-developed chip Jalapeño) and Anthropic are actively developing customized chips, and the AI industry is accelerating towards a dual-track development stage of "model + chip" hardware and software integration.
Anthropic has launched its self-developed AI chip and is in talks with Samsung regarding a customization plan.
According to Beating's monitoring, Anthropic is in talks with Samsung regarding a custom AI chip and has already begun early development work on its self-developed AI chip. Anthropic has also held talks with several chip design companies. --------------------------------- Click the original link below to join Beating's AI news channel on Lark, and receive 24/7 monitoring of global AI hot topics and news.
ByteDance is expected to complete the design of its next-generation self-developed CPU as early as the beginning of next year, with the goal of mass production and large-scale deployment in the second half of next year.
According to a report by the South China Morning Post on June 29th, three sources familiar with the matter revealed that ByteDance plans to complete the design of its next-generation self-developed CPU by early next year at the latest, aiming for mass production and large-scale deployment in the second half of 2027 to support its expanding artificial intelligence footprint. One source indicated that an early version of the self-developed CPU has been in use internally since the end of last year. However, another source stated that given the urgent demand, the tape-out timeline for the new generation CPU may be brought forward. For ByteDance, this CPU is part of its strategy to build its own chip portfolio and integrate more self-developed hardware into its AI infrastructure. Two sources revealed that to accelerate development and assist in securing wafer fab capacity, ByteDance is collaborating with Qualcomm, a US chipmaker actively expanding its AI data center business. This collaboration with Qualcomm, which uses a fabless model and outsources manufacturing to foundries like TSMC, aims to help ByteDance cope with the highly strained advanced chip supply chain. --------------------------------- Click the original link below to join the Beating · Lark AI news channel and monitor global AI hot topics and news 24/7.
BYD's self-developed intelligent driving chip is expected to be first installed in mass-produced vehicles in 2027.
BYD plans to debut its self-developed intelligent driving chip, the Xuanji A3, in its new Denza brand vehicles next year. Official information shows that this 4nm chip boasts a single-chip computing power exceeding 700 TOPS, with three chips working together achieving a total computing power exceeding 2100 TOPS. It supports L3 and L4 autonomous driving and has already entered mass production. The "integrated hardware and software" advantage, commonly emphasized by emerging electric vehicle manufacturers, was also mentioned by BYD during the chip's release. BYD claims that the Xuanji A3's power consumption per unit of computing power is 20% lower than comparable products, and combined with self-developed algorithm optimization, its computing power utilization rate is increased by 100%. (LatePost)
Meituan open-sourced its trillion-parameter large-scale model LongCat-2.0, and simultaneously released the inference code for domestically developed Chinese card processors.
According to Beating's monitoring, Meituan has officially open-sourced its trillion-parameter large-scale model, LongCat-2.0, with a total of 1.6T parameters and an average activation of approximately 48B, designed specifically for real-world agentic coding tasks. Architecturally, it innovatively introduces LongCat sparse attention and N-gram embedding. The former reduces fragmented memory access through flow-aware indexing and hierarchical indexing, accelerating training and inference with millions of contexts; the latter, while achieving nearly 97% sparsity in MoE, invests 135B parameters into the embedding layer, balancing parameter gains and structural stability. Post-training employs multi-teacher online distillation, categorizing experts into Agent, Inference, and Interaction types, seamlessly integrating them on a domestic computing power cluster through the MOPD architecture. As the industry's first trillion-parameter model to complete inference on a 50,000-card domestic computing power cluster, LongCat-2.0 validates the mature capability of domestic chips to handle complex large-scale model tasks. To address the multiple limitations of domestically produced Chinese chips in terms of memory, bandwidth, and interconnects, Meituan has made breakthroughs in three areas: model, chip adaptation, and deployment. At the model level, ScMoE leverages the core control capabilities of domestically produced chips to achieve physical core-level parallelism for Dense and MoE branches, combined with KV-cache partitioning to alleviate the pressure on ultra-long context memory. At the chip adaptation level, Super Kernel reduces operator startup overhead, and Weight Prefetch hides I/O latency, maximizing hardware utilization under constrained conditions. At the deployment level, PD separation is adopted to balance TTFT and TPOT, along with asynchronous Expert-Parallel load balancing to solve load unevenness under high EP (efficiency level). This open-source release simultaneously provides multiple precision versions, including BF16, FP8, and INT8, and fully opens up inference results optimized for domestic computing power, aiming to enable existing domestically produced cards and even older cards to smoothly deploy trillion-model inference services. --------------------------------- Click the original link below to join the Beating · Lark AI news channel and monitor global AI hot topics and news 24/7.
Vibe Coding platform Base44 launched its first self-developed model, Base1, trained with tens of millions of real development records to provide "product intuition".
According to Beating, Wix's Vibe Coding platform, Base44, announced the launch of its first self-developed large-scale model, Base 1, which is already live in production. Unlike platforms that only call general model interfaces, Base44 attempts to achieve full-stack vertical integration from databases and backend infrastructure to the intelligence layer through its self-developed model. The core advantage of Base 1 lies in its unique data training source: tens of millions of real user development records accumulated by the platform in the past. This includes users' original requirements, code generated by the agent, errors during operation, modifications made by users, and real closed-loop signals of whether the application ultimately runs successfully. Base44 founder Maor Shlomo pointed out that general-purpose large-scale models must take into account various programming languages and a wide range of tasks, while Base 1 focuses on specific application development environments. They not only train the model to write code, but also train it to have "product decision-making ability," enabling it to participate in product discussions, predict users' next requirements, and even avoid unreasonable development paths. Currently, Base 1 has been added to the platform's model selector, alongside other mainstream models, and the system will automatically schedule it according to the task type. Base44 plans to release specific technical details and benchmark results for the model in the coming weeks.