Approaching.AI Raises 1.3 Billion Yuan as Chinese Firms Push Cheaper Inference and Physical AI

The company said its core operating metrics grew more than tenfold in half a year and that daily output of high-quality AI tokens rose above one trillion. It has crossed its cost line and is building or operating more than ten token-factory projects with daily trillion-token capacity, including a planned 10,000-GPU-level facility in Hangzhou. The funds will go to capacity expansion, upgrades to its ATaaS platform, inference R&D and commercial delivery, QbitAI reported.

Inference efficiency is also the target of two open-source releases. Tsinghua University and CUHK researchers open-sourced AKV, a serving system for KV cache engineering in LLM agents on SGLang, MarkTechPost reported. AKV lets agents drop selected KV cache entries and reuse the rest; across three models and four agent benchmarks, it cut average cost by 27.7% and raised average task score by 1.7 points. RockMatrix, affiliated with Yanshan Technology, together with HKUST and CMU, open-sourced EdgeRock-4B, which keeps KV cache at 72 MiB at 256K context, a 99.1% reduction, and runs 512 concurrent long-reasoning requests on one 80GB GPU at 12,124 tokens/s, or 6.42 times the source model, according to QbitAI.

In physical AI, DeepMirror presented a full-stack steel-industry inspection system at the 2026 Unitree Technology Partner Ecosystem Conference in Hangzhou on Oct 22, QbitAI reported. The system uses Unitree quadruped robots with DeepMirror’s Insight spatial cameras, Navcore spatial-intelligence brain and task-management platform. It maps plants without pre-placed QR codes or magnetic strips, reroutes around temporary obstacles, carries visible-light, infrared, acoustic and gas sensors, and archives inspection data. The system has entered routine operation at Xiangtan Steel and Jinnan Steel, according to the report. Unitree founder Wang Xingxing said the biggest hurdle to embodied intelligence reaching a ChatGPT moment is aligning AI model inputs and outputs with real robots.

Shenzhen-based Primotion, founded three months ago, unveiled its first public result, the Paimon-0 embodied foundation model, QbitAI reported. The company said Paimon-0 ranked first on RoboColiseum with 0.7754 points, RoboCasa365 with 58.1 and VLA-Arena with a 0.73 average success rate, and won the IROS Workshop challenge with 46.599 points. A customer acceptance record showed that after replacing a software control scheme with Primotion’s FPGA-based one, the worst arrival error at a 1 kHz control cycle fell from 30 microseconds to 152 nanoseconds. Primotion’s founders include Ren Qingyuan, former vice president of a RISC-V chip company, and co-founders Ma Zhiheng and Chang Xinyuan, who have worked in computer vision and embodied intelligence.

iFlytek released Spark-X2.6, a 293-billion-parameter MoE model with about 30 billion active parameters and 1M context, QbitAI reported. The model is trained and served on domestic computing platforms; iFlytek said training and inference architecture optimization raised efficiency on domestic training platforms fourfold and that about 60% of tasks in high-frequency scenarios consumed fewer tokens. The API is priced at 0.9 yuan per million input tokens and 3.6 yuan per million output tokens. iFlytek also open-sourced a 4B version that can run offline on local hardware; in a test, it completed a business review task in 6.8 minutes on CPU only after the network was disconnected.

Editor's Summary

Approaching.AI’s Series B highlights continued capital flow into Chinese AI inference infrastructure. Open-source work from Tsinghua and RockMatrix, along with new embodied models and domestic-compute deployments, shows cost and hardware efficiency are becoming central competitive fronts.