Aldehir Rojas

Open-source work

Aldehir Rojas is listed in the ggml-org/llama-common team in llama.cpp's CODEOWNERS file, the group that owns the project's shared common code. llama.cpp is distributed under the MIT license.¹⁴

Much of Rojas's recent work on llama.cpp concerns chat templates and tool calling. Commits from mid-2026 add specialized parsers for Qwen3 and MiniCPM5, tighten function-call parsing for Qwen models, fix tool-call handling for LFM2 and Gemma 4, and refactor the chat API; an October 2026 commit implements a unique filter in the project's Jinja template engine.²

Rojas has also worked on grammar generation and sampling, including an internal representation for JSON schemas, spacing rules aligned between JSON-schema grammars and parsers, limits on repetition in grammars, and support for multiple end sequences in the reasoning-budget sampler. Other commits change how llama-server splits prompt batches and handles checkpoints.²

Products and projects

Rojas publishes gpt-oss-adapter, an MIT-licensed middleware proxy written in Go that sits between client applications and inference servers such as LM Studio and llama.cpp. It caches reasoning content from tool-call responses for gpt-oss models and injects it into later requests, because, as the README explains, most clients do not send that context back.³

Work and contributions

• llama.cpp Open-source code

  CODEOWNERS lists Rojas in the ggml-org/llama-common team, which owns common/. Commits authored from June to October 2026 add specialized chat parsers (Qwen3, MiniMax M3, MiniCPM5), fix tool-call grammars, implement an internal representation for JSON schemas, extend the reasoning-budget sampler, refactor the chat API, and change llama-server checkpoint and batch handling. The repository is MIT-licensed.¹²⁴

• gpt-oss-adapter Open-source code

  Published under Rojas's GitHub account as an MIT-licensed middleware proxy that manages chain-of-thought reasoning for gpt-oss models, with provider support for LM Studio and llama.cpp.³

Timeline

1. Jun 1, 2026Added support for manually triggering the reasoning-budget end sequence in llama.cpp.²
2. Jun 28, 2026Implemented a MiniCPM5 chat parser in llama.cpp with a co-author.²
3. Jul 25, 2026Added support for multiple end sequences in llama.cpp's reasoning-budget sampler.²
4. Aug 2, 2026Added a specialized Qwen3 chat parser to llama.cpp with a co-author.²
5. Sep 12, 2026Implemented an internal representation for JSON schemas in llama.cpp's common code.²
6. Oct 9, 2026Refactored the llama.cpp chat API.²

Where they work

• llama.cpp project 12 people in the wiki

Related people

• Sigbjørn SkjæretAlso at llama.cpp project
• Adrien GallouëtAlso at llama.cpp project
• Daniel BeveniusAlso at llama.cpp project
• Piotr WilkinAlso at llama.cpp project
• Radoslav GerganovAlso at llama.cpp project
• Dane KnechtInference and serving · Agents
• Philipp SchmidDeveloper tools · Agents
• Olivier ChafikDeveloper tools · Agents
• Amjad MasadDeveloper tools · Agents
• Ben DavisDeveloper tools · Agents