kimodo.cpp

Kimi K3 in C is a portable C99 inference engine for running the 2.78-trillion-parameter Kimi K3 base model on CPUs without BLAS, an inference framework, or a GPU. It reads the 1.56 TB checkpoint from disk, keeps a selected depth of the 93-layer dense trunk in memory, and streams the remainder during generation; its routed experts are packed in 4-bit form, multiplied directly from that representation, and managed with an LRU cache. The implementation includes KDA attention with a non-growing memory, MLA attention with a latent representation, and selection of 16 experts from 896. The repository provides a command-line interface, weightless and full-checkpoint validation gates, and measurement data; it reports operation from 8 GB of RAM upward, with additional memory changing speed rather than the generated output. The base model has no chat template, so prompts produce continuations rather than chat replies. Reference requirements include an AVX2-and-FMA CPU, about 1.7 TB of free storage, and a C compiler; Linux/x86-64 is the reference platform, while macOS/arm64 and Windows/x86-64 are also documented as building and passing the tests.

Kimchi is an open-source coding-agent platform and harness developed by Cast AI. Its terminal coding-agent CLI, built on the pi-mono coding-agent SDK, connects to Kimchi's LLM infrastructure and can operate in single-model or multi-model mode. In multi-model mode, an orchestrator classifies work into roles such as planning, building, reviewing, codebase exploration, and research, then delegates tasks to configured models or model pools using tier and task metadata; requests are tagged by phase and user-defined labels for usage and cost attribution. Kimchi includes Ferment, a persistent project-management mode that stores goals, phases, steps, decisions, memories, and append-only transition logs so work can resume across sessions. Ferment V2 provides an experimental objective controller with token budgets and independent completion evaluation. The CLI also provides Language Server Protocol tools for diagnostics, navigation, references, renaming, and type information, plus remote sessions that synchronize a working tree to a cloud sandbox and keep an agent running after the local terminal closes. Kimchi supports MCP and native Pi packages, migration of configurations from Claude Code, OpenCode, and Cursor, configurable command hooks, and installation through Homebrew, shell or PowerShell installers, or standalone binaries. The repository is licensed under Apache License 2.0.

kimi-k3-mlx is an MLX port for Apple Silicon of Moonshot’s Kimi K3 native-multimodal mixture-of-experts model. It provides a stock `mlx-lm`-compatible text model definition, a 3D video-capable MoonViT vision tower with a patch-merger projector, and an `mlx-vlm`-shaped wrapper that joins the language and vision components. The model has a 1-million-token context and uses routed and shared experts, Kimi Delta Attention, gated multi-head latent attention, Attention Residuals, latent-dimension experts, and the SiTU-GLU activation; routed experts are distributed in MXFP4 while other weights use bf16. The multimodal processor expands each image placeholder into a feature block rather than performing a same-length one-token-per-image scatter. The repository includes a streaming converter because standard `mlx-lm` conversion would materialize the multi-terabyte bf16 model, along with REAP expert-pruning scripts and expert-overlap analysis using Chinese, English, code, or mixed calibration data. It also contains vision-parity tests against Moonshot’s reference implementation. The repository states that the full model cannot run on a single Mac: even its smallest deployment tier requires hundreds of gigabytes of memory and exceeds the capacity of the largest Apple Silicon machine.

Kimi Code CLI is an AI coding agent developed by Moonshot AI that runs in a terminal. It reads and edits code, runs shell commands, searches files, fetches web pages, and selects subsequent actions based on the feedback from those operations. It works with Moonshot AI's Kimi models and can be configured with other compatible model providers. The tool is distributed as a single binary and provides an interactive terminal UI. It accepts video input, supports conversational configuration of Model Context Protocol (MCP) servers, and can install skills, MCP servers, and data sources from its marketplace or GitHub repositories. Built-in coder, explore, and plan subagents can work in isolated contexts, while lifecycle hooks can run local commands at selected points in a session. Kimi Code CLI also supports the Agent Client Protocol (ACP), allowing compatible editors and IDEs such as Zed and JetBrains to drive sessions through the `kimi acp` command. The repository documents installation for macOS, Linux, and Windows and provides OAuth or Moonshot AI Open Platform API-key login options.

BongoCat is a native animated desktop companion that reacts to keyboard, pointer, and gamepad input. It is built in C and C++ with SDL3 and OpenGL, supports Windows, macOS, and Linux, and processes keyboard and mouse input locally rather than recording or uploading it. The runtime uses platform-specific input listeners, an atomic input queue, an SDL3 main loop, model-parameter updates, and OpenGL composition for the pet window and overlays. It includes built-in model assets and supports external model sources, including Mver packages, Tauri sources, `.model3.json` files, and image patches. Live2D rendering uses the Cubism SDK when available; without it, the project builds a diagnostic backend. The source code and native runtime are licensed under AGPL-3.0-only, while the default model mode and bundled model assets have separate MIT licensing.