3 AI Open-Source Projects: DeepSeek Agent Runtime, Verified Diagrams, 744B MoE on 25GB RAM

Spend 3 minutes a day to master the most noteworthy AI open-source projects on GitHub. Today's three projects form a complete pipeline: DeepSeek open-sourced the Agent runtime layer, archify stops AI from hallucinating architecture diagrams, and colibri lets your old computer run cutting-edge large models. Each is more hardcore than the last — 3 minutes to read them all. Are you ready?

1. deepseek-harness ⭐ 230.6k

One-sentence summary: DeepSeek's officially open-sourced Agent runtime framework where everything is a plugin.

Core Features

Everything-is-a-plugin architecture: Model adapters, tool registration, session logging, and even the Agent loop itself are plugins. The foundation is the Cordis composable framework, backed by an arXiv paper.

One-command launch: npx @deepseek-ai/dsh web starts the Web UI on local port 3080; just enter an API key to use.

Rapid ecosystem growth: Media reports indicate the plugin catalog collected 1000+ plugins within two to three days of open-sourcing. Models are not locked to DeepSeek; Anthropic, OpenAI, Azure, and others are supported.

Use Cases

Developers who want to build custom Agents on a unified plugin architecture; teams needing an extensible runtime rather than a single chat assistant.

Verdict

The ambition is huge — it peels open the closed 'black box' layer of tools like Claude Code and Codex, making every component swappable. However, the official repository still carries a 'developer preview' warning about breaking changes. Recommendation: study the architecture first, don't rush to production.

GitHub: https://github.com/deepseek-ai/deepseek-harness

2. archify ⭐ 67.7k

One-sentence summary: Lets Cursor and Claude Code generate interactive architecture diagrams directly in chat.

Core Features

Generates 5 diagram types in chat: Architecture, workflow, sequence, data flow, and lifecycle diagrams. Describe the system and it draws — no repository connection needed.

Deterministic compilation: The model only outputs structured JSON; archify handles compilation and rendering. Topology cannot hallucinate; all checks must pass before delivery, with fix suggestions on failure.

Architecture Delta: Compares base vs head before merge, precisely listing added, removed, changed, and moved components. Understand architectural changes at a glance during PR review.

Use Cases

Developers who want to quickly visualize systems as architecture diagrams within coding agents; teams needing to compare architectural changes before PR reviews.

Verdict

Equips coding agents with a 'diagram tool that doesn't lie.' Previously AI diagrams relied on model guesswork; now the model emits JSON and the program renders, making results verifiable and reproducible. Worth installing for anyone doing technical documentation or PR reviews.

GitHub: https://github.com/tt-a1i/archify

3. colibri ⭐ 36.4k

One-sentence summary: A pure C zero-dependency inference engine that runs 744B to 2.8T MoE models on consumer hardware.

Core Features

Disk-streaming for massive models: Mixture-of-Experts (MoE) models activate only ~5.4% of parameters per token. colibri offloads inactive experts to disk and streams them on demand, enabling a 744B model to run in 25GB RAM.

Zero-dependency pure C: No BLAS, no GPU, no Python runtime required. The Vulkan backend even runs on abandoned older AMD GPUs.

Semantic hard guarantees: Insufficient memory only slows execution; it never silently changes precision or routing. The project performs token-level exact verification for every supported model.

Use Cases

Developers who want to run frontier MoE models locally on consumer hardware, and teams researching inference systems.

Verdict

It is genuinely slow — cold start throughput is only 0.05 tokens/sec on typical machines, nearly 'runnable but not usable.' But it is genuinely correct: it thoroughly solves the pain point of 'touching frontier large models without high-end GPUs.' Suitable for research, not for daily use.

GitHub: https://github.com/JustVugg/colibri

Summary

The three projects each cover a different direction: Agent runtime, developer tooling, and inference infrastructure. A notable commonality is that they all tackle problems 'beyond the model' — AI competition is shifting from model-centric to runtime and tooling.

deepseek-harness (Agent Framework) — ⭐⭐⭐⭐ — Official release with arXiv-backed architecture, but still in developer preview.

archify (AI Coding Tool) — ⭐⭐⭐⭐ — Deterministic, hallucination-free diagrams that directly boost productivity in documentation and review.

colibri (Inference Engine) — ⭐⭐⭐⭐ — Hardcore engineering sample with maximal research value; daily use can wait.