ModelBest's 'Open Source Week': A Systemic Declaration Defining the Endgame of On-Device AI
From May 25 to 29, ModelBest jointly organized an 'On-Device LLM Open Source Week' with the OpenBMB community, releasing five key technological achievements that form a full-stack closed loop: BitCPM-CANN (1.58-bit low-bit training model supporting Ascend), MiniCPM5-1B (outperforming models twice its size), ForgeTrain (AI-written training framework 10% faster than Megatron), PilotDeck (agent operating system), and UltraData (core dataset). These releases demonstrate that the on-device AI competition is a systemic engineering challenge, not a single technology race. MiniCPM5-1B surpasses parts of GPT-4o, validating the 'density law.' ModelBest's two-year lead and deep tech stack position it as a key player in the shift from cloud to edge.
Article intelligence
Key points
- ModelBest held an On-Device LLM Open Source Week from May 25-29, 2026, releasing one key technology each day.
- The five releases cover training framework, model compression, data, and agent OS, showcasing systemic innovation.
- MiniCPM5-1B exceeds GPT-4o on some benchmarks, confirming the 'density law' that smaller models can match larger ones.
- On-device AI will reshape value chains and business models; ModelBest's two-year head start gives it a systemic advantage.
Why it matters
This matters because modelBest held an On-Device LLM Open Source Week from May 25-29, 2026, releasing one key technology each day.
Technical impact
May affect model selection, inference cost, product capability, and evaluation benchmarks.
From May 25 to 29, 2026, ModelBest, in collaboration with the OpenBMB open-source community, held an 'On-Device LLM Open Source Week,' releasing one key technological achievement each day. This rare collective move in the global AI landscape echoes DeepSeek's similar event in June 2024, but with a distinct focus on edge computing.
The five released technologies form an interconnected system rather than isolated breakthroughs. They are: BitCPM-CANN, a 1.58-bit low-bit training model adapted for China's Ascend hardware, potentially enabling 60-billion-parameter models to run on phones; MiniCPM5-1B, a 1-billion-parameter model that outperforms models twice its size and rivals GPT-4o on certain tasks; ForgeTrain, an AI-written production-grade training framework that is 10% faster than Nvidia's Megatron on H100 GPUs; PilotDeck, an agent operating system that redefines interaction paradigms; and UltraData, a core dataset that reveals the source of on-device models' efficiency.
These releases underscore that the endgame of on-device AI is not about a single standout technology but about systemic innovation across data, algorithms, frameworks, and applications. ModelBest's approach is rooted in genuine open source—it not only opens the model but also the 'production line' (ForgeTrain), 'core process' (BitCPM), 'raw materials' (UltraData), and even the 'agent OS' (PilotDeck). This level of openness is rare globally and reflects a deep technical conviction.
ModelBest's journey in open source began early. The OpenBMB community, co-founded by Tsinghua University's THUNLP lab and ModelBest in 2022, started China's earliest full-chain LLM open-source exploration. It offered free public courses on large models and has accumulated over 130,000 GitHub stars, ranking among the top 100 global open-source organizations. Its MiniCPM series has been downloaded over 30 million times, and UltraData over 4 million times.
The company's systemic advantage is evident. ForgeTrain, for instance, is entirely written by AI with zero human code—a milestone that moves development into an 'AI manufacturing AI' phase (L3+). The full-stack loop covers infrastructure (BMTrain, BitCPM), data curation (UltraData), model algorithms (MiniCPM, VoxCPM), and applications (PilotDeck, smart cockpit, legal AI). This systemic synergy is hard to replicate because on-device AI requires delicate balance across hardware, software, algorithms, and data—unlike cloud AI where scale can brute-force results.
ModelBest's two-year head start in on-device AI (since 2024) gives it a formidable moat. As other players rush into the space in 2026, ModelBest has already refined its technology stack. MiniCPM5-1B's performance exceeding GPT-4o in some areas confirms the 'density law': on-device models are rapidly encroaching on cloud model territory.
The implications for the industry are profound: value will shift from cloud API providers (OpenAI, Google) to hardware makers and on-device AI OS controllers; applications will enjoy offline availability, zero latency, and absolute privacy as default; and business models may move from per-token billing to software licenses or hardware bundling. This represents a tectonic shift in AI's power and value chain.
Comparing ModelBest's Open Source Week with DeepSeek's reveals two complementary visions for Chinese AI: DeepSeek pushes the limits of cloud model capabilities with efficiency aesthetics, while ModelBest drives widespread deployment of powerful AI to every edge device. Together, they form a yin-yang of China's AI competitiveness.
Ultimately, the path to AGI involves making intelligence as ubiquitous as air and water. ModelBest, with its systemic approach and two-year lead, is a torchbearer on this road, releasing AI from cloud servers into everyone's pocket, car, and home.