ds4 vs colibri
Local LLM inference engines compared
ds4 and colibri are C-based engines for running open-weight language models locally. ds4 focuses on selected model families and hardware-specific workflows, while colibri centers on streaming Mixture-of-Experts weights from disk to reduce memory demands.
ds4: Run DeepSeek and Other LLMs Locally
ds4, also called DwarfStar, is a native inference engine for selected open-weight language models on Apple Metal, NVIDIA CUDA, and AMD ROCm. It suits users with capable local hardware who want model-specific inference, serving, or coding-agent workflows.
colibri: Run Large MoE Models on Local Hardware
colibri is a C inference engine that streams mixture-of-experts model weights from disk, letting large open models run on machines with limited RAM and no required GPU. It also provides a local dashboard and OpenAI- and Anthropic-compatible APIs.
| ds4 | colibri | |
|---|---|---|
| Language | C | C |
| License | MIT | Apache-2.0 |
| Stars | 23.2k | 40k |
| Forks | 2.2k | 4.4k |
| Last analyzed | Oct 3, 2026 | Oct 6, 2026 |
Key differences
- ds4 targets selected model families, including DeepSeek V4, GLM, and Qwen; colibri supports several MoE families, including GLM, Qwen, Kimi, DeepSeek, Inkling, and OLMoE.
- ds4 provides model-specific inference, with SSD streaming and distributed execution options for supported setups; colibri streams routed MoE experts from disk using caching, placement, and optional prefetching.
- ds4 supports Apple Metal, NVIDIA CUDA, and supported AMD ROCm; colibri supports CPU inference and optional CUDA, Metal, and Vulkan backends.
- ds4 bundles an interactive CLI, local HTTP server, and native coding agent; colibri offers chat, serve, and web entry points, an OpenAI-compatible API gateway, and a browser dashboard.
- ds4 is licensed under MIT; colibri is licensed under Apache-2.0.
- Both projects are written in C, but colibri uses Python for its launcher, gateway, and some tools. ds4 describes itself as beta quality, while colibri presents itself as an open research platform with some experimental or workload-dependent components.
Choose ds4 if you…
- want a model-specific engine with a native coding agent and interactive CLI.
- need supported multi-GPU or distributed execution options for selected models.
- plan to run supported models on Apple Metal, NVIDIA CUDA, or AMD ROCm hardware.
Choose colibri if you…
- need to run MoE models whose full expert set does not fit in RAM or VRAM, and can accept disk I/O tradeoffs.
- want to experiment with expert placement, caching, storage I/O, or CPU and GPU execution.
- prefer an included web interface or OpenAI-compatible API gateway, or need Brio scoring for fixed answer choices.
Compare the numbers side by side, or add more projects
This comparison is generated with AI from the OSRepos analyses of both projects. Always check each project's repository and documentation before choosing.