Piotr Wilkin

Open-source work

Wilkin's contributions to llama.cpp began in May 2025 with a fix to the server's handling of OpenAI-compatible message arrays, and soon extended to adding model architectures such as Ernie 4.5 MoE (July 2025) and Apertus (October 2025). In December 2025 Wilkin was added to the project's CODEOWNERS for its chat files, and later work added Qwen3.5 dense and mixture-of-experts support (February 2026) and the Kimi-K3 text model (August 2026).^67891011

Much of Wilkin's later llama.cpp work concerns the shared parsing code that turns model output into reasoning text and tool calls. In September 2026 Wilkin split the specialised chat parsers into their own common/parsers module, and the current CODEOWNERS file lists Wilkin for tools/parser and the chat tests and as part of the llama-common team.¹²²³

Products and projects

Wilkin writes standalone C++ runtimes built on GGML. trellis.cpp ports Microsoft's TRELLIS.2 image-to-3D pipeline to native C++ with CUDA, ROCm and Vulkan builds, and thinksound.cpp runs the ThinkSound text-to-sound-effect model without Python at inference time. trellis.cpp serves as the trellis backend of Lemonade, where Wilkin's July 2026 commits added asset generation with Trellis and OpenMOSS and audio generation with ThinkSound and ACE-Step.¹³¹⁴⁴

Earlier tools from Wilkin include Catapult, a desktop and terminal launcher that manages llama.cpp runtimes, models and server settings, whose README now points users to Lemonade instead.¹⁵

Work and contributions

• llama.cpp Open-source code

  CODEOWNERS names Wilkin for tools/parser and tests/test-chat, co-owner (with Daniel Bevenius) of examples/debug, and a member of the ggml-org/llama-common team. Commits authored from June to October 2026 split specialized chat parsers into common/parsers, fix DeepSeek and other chat templates and tool calls, add model support such as Kimi-K3, and tune CUDA, HIP and Vulkan kernels. The repository is MIT-licensed.²³¹⁶

• Lemonade Open-source code

  Commits authored from July to September 2026 add asset generation (Trellis and OpenMOSS), audio generation backends (ThinkSound and ACE-Step) and a job engine, and fix how Lemonade finds and launches llama.cpp. The repository is Apache-2.0 licensed and states that the project is sponsored by AMD.⁴⁵

Timeline

1. May 15, 2025First llama.cpp contribution merged, fixing error handling for OpenAI-compatible message arrays.⁶
2. Jul 17, 2025Added Ernie 4.5 MoE model support to llama.cpp.⁷
3. Oct 2, 2025Added the Apertus model implementation to llama.cpp.⁸
4. Dec 5, 2025Added to llama.cpp's CODEOWNERS for the chat files.⁹
5. Feb 8, 2026Added Qwen3.5 dense and mixture-of-experts support to llama.cpp.¹⁰
6. Jun 18, 2026Made the initial commit of trellis.cpp, a C++/GGML port of the TRELLIS.2 image-to-3D pipeline.¹³¹⁷
7. Jul 2026Added asset- and audio-generation backends and a job engine to Lemonade.⁴
8. Sep 8, 2026Split llama.cpp's specialised chat parsers into a common/parsers module.¹²

Where they work

• Syndatis 1 person in the wiki
• llama.cpp project 12 people in the wiki

Related people

• Adrien GallouëtInference and serving · Developer tools
• Tim DavisDeveloper tools · Inference and serving
• Akshat BubnaInference and serving · Developer tools
• Matt HicksInference and serving · Developer tools
• Pierrick HymbertInference and serving · Developer tools
• Jeffrey MorganInference and serving · Developer tools
• Julien SimonInference and serving · Developer tools
• Yagil BurowskiInference and serving · Developer tools
• Matthew PrinceInference and serving · Developer tools
• Luis Héctor ChávezDeveloper tools · Inference and serving