Horizon Robotics plans to supply chips and driving assistance system for BYD's God's Eye B package

The Chinese startup Apex Intelligence, founded in June 2026, unveiled its first model, Apex-35B-V1.0, on October 28. The model weights are available on the Hugging Face platform. The model is based on Alibaba's Qwen3.6-35B-A3B and was further refined through intermediate and subsequent training focused on coding tasks, agent workflows, general tasks, as well as research in artificial intelligence, computational science, and robotics simulation.

The company's main focus is not so much on the model's size as it is on its creation methods. Apex Intelligence asserts that its automated research system, named Apex Research, participated in generating training data, conducting experiments, developing the training plan, and optimizing GPU cores and the inference system.

As an example, the company cited a case where Apex Research discovered a weakness in the model's computational fluid dynamics tasks. The system analyzed the reasons for the model's shortcomings in reasoning, then generated and verified new training data, which was used in subsequent training. According to the company's internal testing, this cycle increased the model's CFD-Bench score by 32.1 points. Furthermore, Apex Intelligence reports that AI-driven optimization increased training throughput by 45% to 60% and accelerated decoding by 1.6–1.8 times.

All benchmarking results published with the release were obtained through the company's own evaluation. Based on this data, Apex Intelligence claims that Apex-35B-V1.0 surpasses both Qwen3.6-35B-A3B and Flint from Recursive SuperIntelligence across all six presented research benchmarks. The model also performs better than GLM-5.3 on FrontierScience and PaperBench tests, outperforms DeepSeek-V4.1-Flash-1017 in the human-difficult BioMysteryBench division, and demonstrates the highest BFCL V4 score among comparable models.

In two robotics tests, RoboRSI and Real2Sim, the model achieved scores of 44.5 and 21.55, respectively, which are close to the performance of the larger GLM-5.3-Flash model (52.4 and 21.4).

The Real2Sim test requires the model to reconstruct a working physical simulation from a single photograph: it must identify objects, select simulation resources with similar visual and physical characteristics, and recreate their arrangement for robot action testing. RoboRSI assesses the model's ability to control simulated robots by writing and correcting code based on visual feedback.

The company was founded by Chen Yunchao, an associate professor at Tsinghua University's AI College. He describes this work as recursive self-improvement—a concept where AI systems help create their successors. In an interview with Tencent Technology, he noted that AI has made clear progress in conducting research, but it still lacks 'research intuition,' meaning the ability to determine which problems are worth pursuing.

Lenovo demonstrated the preliminary model of its AI Host P7 device at the CSDN's 2026 1024 Programmers' Day and AI Builders Conference on October 24. Additionally, an early access program was launched, inviting initial users to test the device in real coding and office tasks before its general release. Previously, the concept of this portable AI host was presented by Lenovo in May, and this preliminary version includes more detailed technical specifications and software features.

The device weighs about 300 grams and is built on two domestic chips. The first is the in-memory computing inference chip Houmo AI M50, which, according to Houmo, provides 160 TOPS performance while consuming 10 watts. The second chip is the agent SoC from CIX Technology P1, manufactured using a 6nm process and equipped with a 12-core Armv9.2 processor and a 30 TOPS NPU, enabling 45 TOPS in heterogeneous computing.

According to CIX, the P1 chip is responsible for system management, task execution, and planning, and it can interact with external accelerators to run a 122-billion parameter model with a 128K context at a speed of 50 tokens per second. Lenovo claims that the preliminary P7 version is capable of locally processing models with up to 125 billion parameters and a 256K context.

The device comes with a local model library that includes Qwen3.8-Flash-Next, Qwen3.8-27B, and Qwen3.6-35B-A3B, as well as models for speech recognition, speech synthesis, document processing, and image design. Lenovo reports that its internal AI computing architecture, which uses mixed INT4 and INT8 quantization, achieves an average of 98.3% of the baseline BF16 with full precision across four benchmarks. It should be noted that these figures are company data and have not undergone independent testing.

Lenovo positions the P7 as a personal token hub. The token manager function allows users to direct coding tools and agent applications to local models via an API address, thereby saving routine model calls on the device instead of using paid cloud quotas. The task execution mode operates around the clock, and progress can be tracked from phones, tablets, and PCs. The built-in Tianxi AI has an Auto mode that distributes requests between local and cloud models depending on their complexity.

Sun Ke, Head of AI Hosts, emphasized that true productivity should not start with counting token costs. Lenovo has not yet disclosed the price or release date, only stating that the P7 will be available soon.