llama.cpp adds support for Aleph Alpha's Kolibri-1 MoE model

llama.cpp has merged support for Kolibri-1, a new Mixture-of-Experts (MoE) reasoning model from Aleph Alpha. The model features 78B total parameters with only 3.46B active parameters per token, designed for efficiency in long-context tasks and multilingual reasoning.

Model Specifications

• Architecture: 50-layer MoE with 4:1 SWA:GQA attention, utilizing 384 experts per layer (1 shared, 6 routed).
• Context Length: Supports up to 1,048,576 tokens, with a recommended serving limit of 262,144 tokens for complex tasks.
• Precision: Uses float8_e4m3fn weights in 128×128 blocks with dynamically quantized activations and an FP8 KV cache. Embeddings and norms are in bfloat16.
• Capabilities: Explicit reasoning mode, tool calling, and strong performance in German and English.

Training and Hardware

• Training Data: Pre-trained on 20T tokens (62.5% English, 23.9% German, 13.6% code) with a knowledge cutoff of June 18, 2026.
• Compute: Pre-training utilized 768 NVIDIA B200 GPUs over 21 days, totaling 392k GPU hours.
• Hardware Requirements: Requires approximately 78 GB of memory for FP8 weights. Minimum configuration includes 2× A100 80 GB or 1× H200.

Availability

• License: Apache 2.0.
• Release Date: October 3, 2026.
• Status: GGUF files are pending release on Hugging Face.