Silicate Zero is live — Ring-0 execution at 149.3 tokens/sec
Silicate Zero is now live — the AI kernel operating system runtime running at Ring-0, directly on bare metal. No Linux. No libc. No syscall overhead.
Benchmark
149.3 tokens/sec on AMD EPYC 9354P, CPU-only, Ring-0. That's a 4–15× lead over above-kernel runtimes like llama.cpp and Ollama on the same hardware class.
Full benchmark methodology and comparison table →
Architecture
- LLM-agnostic runtime — Qwen3-1.7B default, any GGUF model is swappable
- Zero syscall overhead — AI is the kernel, not an application running on top of one
- Single address space — model weights, KV cache, tokeniser buffers all live flat in memory
- Direct hardware telemetry — CPU perf counters, thermal sensors, memory-controller stats read natively
Product Family
- Silicate Zero Basic — Open source (AGPLv3), GitHub public. The full Ring-0 runtime.
- Silicate Zero Bare — Commercial license for OEM / embedded. Contact for pricing.
- Silicate Compute — AI execution layer. Tensor scheduling, batch inference, model orchestration.
- Silicate Nefesh — Biometric integration layer. Hardware-level biometric data processing.
- Silicate Cloud — Managed Ring-0 inference. Coming soon.
- Silicate Benchmark — Open-source benchmark harness + managed hardware rental.
What's next
- Silicate Cloud — managed Ring-0 inference as a service
- Silicate Benchmark Managed — dedicated test servers for enterprise
- Silicate Vault — encrypted model storage at Ring-0