120B
- Context
- 1Mtokens
- Released
- 1mo ago
- Size
- 78.1Bparameters
Aleph Alpha's Kolibri 1, built on the base model Aleph-Alpha/Kolibri-1-BF16, is the company's mixture-of-experts reasoning model focused on German and English. Its maker says the design delivers strong bilingual performance at low serving cost, activating only a small fraction of parameters per token and using a tokenizer tailored to German word structure. It is intended for multi-step reasoning, coding, structured extraction, retrieval-augmented generation, long-document processing, and agentic tool calling in human-reviewed workflows.
- Language model
- Open weights · 78.1B
- 1M context
- Reasoning
- Tools
How it compares
Additional source evidence
- Source model
- aleph-alpha/kolibri-1
- Last updated
- 2026-10-03
- Knowledge cutoff
- 2026-06-18
- Model context
- 1M
- Open weights
- Yes
- File attachments
- No
- Reasoning
- Yes
- Tool calling
- Yes
- Temperature control
- Yes
Downloadable weights
Complete source record (JSON)
{
"id": "aleph-alpha/kolibri-1",
"name": "Kolibri 1",
"description": "Aleph Alpha's German and English mixture-of-experts reasoning model (78B total, 3.46B active) for chat, RAG and agents",
"last_updated": "2026-10-03",
"knowledge": "2026-06-18",
"attachment": false,
"reasoning": true,
"tool_call": true,
"temperature": true,
"modalities": {
"input": [
"text"
],
"output": [
"text"
]
},
"limit": {
"context": 1048576
},
"release_date": "2026-10-03",
"open_weights": true,
"license": "Apache-2.0",
"weights": [
{
"label": "Hugging Face",
"url": "https://huggingface.co/Aleph-Alpha/Kolibri-1"
}
]
}Frequently asked questions
Running locally
At Q4, Kolibri 1 needs an estimated 55 GB of VRAM, 95.9 GB at Q8 and 180.1 GB at FP16. These are estimates from the 78.1B parameter count, covering weights plus about 15% and 0.5 GB for a short context.
No. Kolibri 1 does not fit on a 24 GB, 16 GB, 12 GB or 8 GB GPU at Q4, Q8 or FP16 (estimate).
Yes. Based on the estimated Q4 requirement of 55 GB, the smallest common Mac memory size that fits is 96 GB, assuming macOS lets the GPU use about 75% of unified memory.
Privacy and data
Yes. Kolibri 1 has open weights, so you can self-host it for full control of your data.
Capabilities and specs
The context window of Kolibri 1 is 1M tokens (1,048,576).
No. Kolibri 1 only supports text as an input modality.
Yes, Kolibri 1 supports tool calling (function calling).
Yes, Kolibri 1 supports a reasoning mode.
The knowledge cutoff of Kolibri 1 is 2026-06-18.
Quality and benchmarks
The closest alternatives are Nemotron 3.5 Super 120B A12B by NVIDIA, ERNIE-5.5-Flash-Preview by Baidu and MiniCPM-V 4.7 1B by OpenBMB. They are matched on model kind and release date, with no comparable price.
Nemotron 3.5 Super 120B A12BKolibri 1 vs Nemotron 3.5 Super 120B A12BERNIE-5.5-Flash-PreviewKolibri 1 vs ERNIE-5.5-Flash-PreviewMiniCPM-V 4.7 1BKolibri 1 vs MiniCPM-V 4.7 1B
Licensing and access
Kolibri 1 is open weights, licensed under Apache-2.0.
Yes. Kolibri 1 uses the Apache 2.0 license, a permissive license that allows commercial use.
You can download the Kolibri 1 weights from the Hugging Face repository Aleph-Alpha/Kolibri-1.
Basics
Kolibri 1 was made by Aleph Alpha and released on October 3, 2026.
Yes. Kolibri 1 is open weights.
Generated from Vertical Tabs data, updated November 12, 2026.
Discussion
No comments yet
Be the first to share your thoughts.