DLLM: Minimal, clean coding agent built directly on llama.cpp without overhead

Danny Arends released DLLM, a minimal coding agent written in the D language that runs directly on llama.cpp via importC, eliminating Python and bindings. The agent operates a three-model pipeline (agent, summary, embed) with CUDA offloading, multimodal vision, RAG, and an extensible tool system, tested with Qwen3.5-4B, Qwen2.5-0.5B, and Nomic embed text v1.5 models, requiring 8GB VRAM.

A minimal, clean D language https://dlang.org/ agent built directly on llama.cpp https://github.com/ggerganov/llama.cpp via importC https://dlang.org/spec/importc.html . No Python, no bindings, no overhead. Runs a three-model pipeline agent , summary , embed with full CUDA https://developer.nvidia.com/cuda/toolkit offloading, multimodal vision via mtmd, RAG, KV-cache condensation, thinking budget, and an extensible tool system auto-registered via user-defined attribute @Tool "Description" on functions . Tools included cover: file I/O, web search, date & time, text encoding, Docker sandboxed code execution, and audio playback. Compilation guide for dependencies are found in deps/README.md D Compiler : Use either DMD, LDC, or GDC to compile the agent. llama.cpp : Bundled with llama.cpp https://github.com/ggerganov/llama.cpp Cuda Toolkit : Built with CUDA https://developer.nvidia.com/cuda/toolkit for GPU acceleration support. SearxNG : SearxNG https://github.com/searxng/searxng is used to provide unlimited web searching by the webSearch tool. Docker : Docker https://www.docker.com/ provides a code execution sandbox used by the runCode tool. Tested with the following HuggingFace https://huggingface.co/ models: Agent : Qwen3.5-4B https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/main/Qwen3.5-4B-Q5 K M.gguf & mmproj-F16 https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/main/mmproj-F16.gguf vision . Summary : Qwen2.5-0.5b instruct https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF/resolve/main/qwen2.5-0.5b-instruct-q4 k m.gguf . Embed : Nomic embed text v1.5 https://huggingface.co/nomic-ai/nomic-embed-text-v1.5-GGUF/resolve/main/nomic-embed-text-v1.5.Q4 K M.gguf . Model paths are configured in src/main.d /DannyArends/DLLM/blob/master/src/main.d and with the above model setup require 8GB VRAM. Execute with prompt to OneShot: dub -- "What is your name?" dub -- "Download the image at https://picsum.photos/400, load it, and write a poem about it" dub -- "What date is it today ? and which day of the week was 2 days ago ?" dub -- "Generate an 8-second audio file of a 440hz sine wave that fades out, then play it" dub -- "Generate a spoken 16-bit PCM WAV of 'Hello World' and play it" dub -- "Ingest the file README.md into RAG, tell me what DLLM is in one sentence" dub -- "Read the file dub.json, and summarize in a single line what the file is about." dub -- "How is the weather in Newcastle upon Tyne UK ?" dub -- "Fetch the Bitcoin price history online, plot it as a chart and save to workspace" dub -- "Think about then tell me, a story about math, 4 lines of text and be creative " Or start an interactive session: dub Want to contribute? Great Contribute to this repo by starring ⭐ or forking 🍴, and feel free to start an issue first to discuss idea's before sending a pull request. You're also welcome to post comments on commits. Written by Danny Arends and released as GPLv3 /DannyArends/DLLM/blob/master/LICENSE.txt