Base Compute
Base Compute is an AI inference lab building inference engines and infrastructure that make powerful AI run on-device. Its mission is to run AGI on-device so people can have fast, private intelligence they own, at near-zero marginal cost.
At a Glance
- Chip vendors and silicon manufacturers
- AI model makers and open-weight model developers
- Device makers shipping on-device AI
- Enterprises with on-premise, hybrid, or air-gapped requirements
- +2 more
AI Tools by Base Compute
(2)
Compute:Arena
Local AI Hardware Benchmark Leaderboard
BaseRT
LLM Inference Runtime for Apple Silicon
Discussions
No discussions yet
Be the first to start a discussion about Base Compute
Latest News
Products & Services
Native Metal LLM inference runtime for Apple Silicon. It uses chip-specific kernel fusion, unified-memory-aware optimization, custom dispatch, and quantization support from Q2 to FP16; it can serve models through an OpenAI-compatible API and supports model families including Qwen, Llama, Gemma, Mistral, Phi, and Nomic BERT.
Private, no-cloud, no-account AI workspace for macOS with on-device chat, file conversations, coding in a local folder, meeting transcription, editable memory, model recommendations, analytics, and optional user-supplied cloud keys.
Optimization stack for chip vendors, model makers, and device makers. It combines per-model quantization, graph rewrites, operator fusion, hand-written and auto-tuned kernels, profiling, and hardware co-design across Apple silicon, NVIDIA, AMD, and other accelerators.
Infrastructure for running AI in on-premise, hybrid, or fully air-gapped environments, keeping models, hardware, data, and policy controls within the customer environment.
Market Position
Base Compute positions BaseRT against llama.cpp and MLX as a faster, native-hardware inference runtime, particularly on Apple Silicon, while B:OS is positioned as a hardware-agnostic model-to-kernel optimization layer for chip, model, and device makers. Its differentiation is low-level, per-chip/per-model optimization, agent-assisted porting, and privacy-preserving local execution rather than cloud API access.
Key Competitors
Leadership
Founders
Lukas Wesemann
Physics PhD and nanotechnology postdoctoral researcher at the University of Melbourne; subsequently worked on the Australian language models Matilda and Maincoder at Maincode.
Prabod Rathnayaka
PhD in artificial intelligence from La Trobe University; worked on Matilda and previously held roles at Replica Studios and John Snow Labs.
Fabian Waschkowski
Mechanical-engineering doctorate focused on evolutionary machine learning and computational fluid dynamics; later moved into inference research.
Executive Team
Prabod Rathnayaka
Chief Technology Officer
Co-founder with a PhD in artificial intelligence from La Trobe University; previously worked at Replica Studios and John Snow Labs and on the Matilda language model.
Founding Story
The three co-founders started Base Compute around the belief that capable models now fit on hardware people and companies already own, making local inference attractive for privacy, latency, offline operation, cost, and independence from cloud providers. Their initial vision was an inference lab working from silicon-level kernels and runtimes upward so people could own and run AGI on their devices.
Business Model
Revenue Model
The company describes B:OS as an offering for chip vendors, model makers, and device makers shipping AI at scale, and describes an enterprise stack for on-premise, hybrid, and air-gapped deployments. Local and BaseRT are publicly downloadable; the public pages do not state subscription or usage pricing.
Target Markets
- Chip vendors and silicon manufacturers
- AI model makers and open-weight model developers
- Device makers shipping on-device AI
- Enterprises with on-premise, hybrid, or air-gapped requirements
- Regulated industries including finance, healthcare, government, and defence
- Developers and engineers running local AI on Apple Silicon and other edge hardware
- Private local chat and document question-answering
- Local coding agents and software development
- On-device meeting transcription and summaries
- Inference for regulated industries with sensitive data
- Offline or low-connectivity AI
- Enterprise fleets of laptops and edge devices