Model : add Kolibri-1 (Aleph-Alpha) support by riteme · Pull Request #30673 · ggml-org/llama.cpp
The model has 78B parameters with 3.46B active per token, featuring 50 layers and 384 routed experts plus one shared expert. Its architecture includes attention mechanisms with sliding windows and a 513-token context, alongside GQA with 48 query heads and 4 KV heads. The converter uses existing AfmoeModel layouts and handles FP8 checkpoints without special code adjustments.
Key points
- Pull request proposes Kolibri-1 support for llama.cpp, a German/English MoE model with 78B parameters
- Model uses 3.46B active parameters per token, 50 layers, and 384 routed experts with sliding attention
- Converter reuses AfmoeModel and handles FP8 checkpoints without special code, but suggestion lacks code changes
The request references DeepSeek Harness (DeepSeek V4.1 Flash) as an assist but notes the suggestion is invalid due to no code changes. The pull request is part of the ggml-org/llama.cpp repository and includes technical details like tensor naming and pre-tokenizer registration.
Model page: Kolibri-1 →
model : add Kolibri-1 (Aleph-Alpha) support by riteme · Pull Request #30673 · ggml-org/llama.cpp
github.com · 17 October 2026Loading the full article…
This text was published by github.com. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Meta releases Muse Spark Open, an 816B-parameter open-weight multimodal model with 42B active parameters and 1M-token context · 32 src
- DeepSeek releases V4.1-Flash-1017 with faster reasoning · 3 src
- TypeSafe AI releases Jev 1.14 with images, box type and 262k token context · 2 src
- StepFun releases Step 5 Preview open weights with 600B parameters and 1M context · 1 src
- OpenAI makes Prompt-to-Voice generally available for developers · 1 src
Comments
via GitHub Discussions