Deltafin

Deltafin runs full Kimi K3 on a single device for researchers and developers exploring large models on their own hardware. It preserves the original expert weights and complete expert bank, but responses are slow and long conversations take substantially longer than short completions.

• Full Kimi K3 inference: K3 determines every output token, including checking suggestions from Inferact's Kimi-K3-DSpark draft model before accepting them.
• Chat and completion API: An OpenAI-compatible server connects local chat clients and coding agents to text chat and raw completions, with streaming responses.
• On-demand model downloads: A storage-saving mode fetches exact experts as they're needed, with much slower initial responses when those experts aren't cached locally.
• Conversation reuse: The server reuses exact conversation state and caches exact responses to reduce repeated processing.

Inference runs on your device. On-demand downloads need internet access for uncached experts. Deltafin has a native Rust runtime, supports NVIDIA/CUDA builds and CPU execution, and has been benchmarked on an Apple M1 Max laptop.

The API supports a limited, text-only subset of OpenAI requests and handles one generation at a time. Output is deterministic, and the server rejects unsupported requests rather than silently ignoring their settings. Network exposure requires your own authentication and access boundary.