{"slug": "dllm-minimal-clean-coding-agent-built-directly-on-llama-cpp-without-overhead", "title": "DLLM: Minimal, clean coding agent built directly on llama.cpp without overhead", "summary": "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.", "body_md": "A minimal, clean [D language](https://dlang.org/) agent built directly on [llama.cpp](https://github.com/ggerganov/llama.cpp) via\n[importC](https://dlang.org/spec/importc.html). No Python, no bindings, no overhead. Runs a three-model\npipeline (*agent*, *summary*, *embed*) with full [CUDA](https://developer.nvidia.com/cuda/toolkit)\noffloading, multimodal vision via mtmd, RAG, KV-cache condensation, thinking budget, and an extensible tool\nsystem (auto-registered via user-defined attribute *@Tool(\"Description\")* on functions). Tools included\ncover: file I/O, web search, date & time, text encoding, Docker sandboxed code execution, and audio playback.\n\nCompilation guide for dependencies are found in `deps/README.md`\n\n**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.\n\nTested with the following [HuggingFace](https://huggingface.co/) models:\n\n**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).\n\nModel paths are configured in [ src/main.d](/DannyArends/DLLM/blob/master/src/main.d) and with the above model setup require 8GB VRAM.\n\nExecute with prompt to OneShot:\n\n```\n  dub -- \"What is your name?\"\n  dub -- \"Download the image at https://picsum.photos/400, load it, and write a poem about it\"\n  dub -- \"What date is it today ? and which day of the week was 2 days ago ?\"\n  dub -- \"Generate an 8-second audio file of a 440hz sine wave that fades out, then play it\"\n  dub -- \"Generate a spoken 16-bit PCM WAV of 'Hello World' and play it\"\n  dub -- \"Ingest the file README.md into RAG, tell me what DLLM is in one sentence\"\n  dub -- \"Read the file dub.json, and summarize in a single line what the file is about.\"\n  dub -- \"How is the weather in Newcastle upon Tyne (UK) ?\"\n  dub -- \"Fetch the Bitcoin price history online, plot it as a chart and save to workspace\"\n  dub -- \"Think about then tell me, a story about math, 4 lines of text and be creative!\"\n```\n\nOr start an interactive session:\n\n```\n  dub\n```\n\nWant 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.\n\nWritten by Danny Arends and released as [GPLv3](/DannyArends/DLLM/blob/master/LICENSE.txt)", "url": "https://wpnews.pro/news/dllm-minimal-clean-coding-agent-built-directly-on-llama-cpp-without-overhead", "canonical_source": "https://github.com/DannyArends/DLLM", "published_at": "2026-08-12 22:34:15+00:00", "updated_at": "2026-08-12 23:11:24.453669+00:00", "lang": "en", "topics": ["ai-agents", "developer-tools", "ai-infrastructure"], "entities": ["Danny Arends", "DLLM", "llama.cpp", "D language", "CUDA", "Qwen3.5-4B", "Qwen2.5-0.5B", "Nomic embed text v1.5"], "alternates": {"html": "https://wpnews.pro/news/dllm-minimal-clean-coding-agent-built-directly-on-llama-cpp-without-overhead", "markdown": "https://wpnews.pro/news/dllm-minimal-clean-coding-agent-built-directly-on-llama-cpp-without-overhead.md", "text": "https://wpnews.pro/news/dllm-minimal-clean-coding-agent-built-directly-on-llama-cpp-without-overhead.txt", "jsonld": "https://wpnews.pro/news/dllm-minimal-clean-coding-agent-built-directly-on-llama-cpp-without-overhead.jsonld"}}