Houmo AI open-sources full software stack for its M50 edge chip
At an ecosystem partner conference on October 14, Houmo AI said it will fully open-source its AI software stack, covering four layers — model quantization and compilation, inference frameworks and operator libraries, domain kits and middleware, and deployment/operations tools — in three batches, with 15 projects due by January, including adaptations of llama.cpp and vLLM; postmo-modelzoo, postmo-quantflow and vllm-postmo are already on GitHub. The company says its M50 chip, less than a year into mass production, has been picked for more than 150 terminal products by over 100 customers, including Lenovo's 300g AI host P7 (up to 48GB LPDDR5/5X, a 122-billion-parameter model running offline at 50 tokens/s), Great Wall's N90 Pro laptop running a 35B model fully offline on Kylin V11, and UGREEN's HomeAgent HA100 Pro. Houmo also claims its SRAM compute-in-memory design cuts chip area by about 70% and power to roughly one-tenth, and that M50 is China's only high-compute edge AI chip compatible with mainstream operating systems and CPU architectures.
Most of the cost of shipping on-device AI sits in adapting and tuning models for a specific chip, so an open toolchain could let smaller teams reuse Houmo's engineering work instead of starting over.