Kolibri-1: install open-weight AI on a 64 GB Mac

Kolibri-1 is a text AI model from Aleph Alpha, based in Heidelberg, Germany. Its weights are available on Hugging Face. A law firm could serve it on its own infrastructure and analyse documents without sending them to an external AI service. Local hosting alone does not guarantee privacy compliance: secure access, retention and review still matter.

Open weights under Apache 2.0 permit commercial use and adaptation in your product. The grant covers published weights and configuration, not the training code. Released on 3 October 2026, this is a German and English text model, not a Portuguese or image model.

Think of a hospital with 384 specialists on each floor: for each word, 6 are called in and 1 shared specialist always participates. This mixture of experts repeats across 50 layers for each token. Of 78.1 billion parameters in total, about 3.46 billion are active per token. You still need memory for all the weights.

python3 -m venv kolibri-env && source kolibri-env/bin/activate

python -m pip install -U mlx-lm huggingface_hub

python -m pip show mlx-lm

hf download velaia/Kolibri-1-MLX-4bit --local-dir Kolibri-1-MLX-4bit

python Kolibri-1-MLX-4bit/run.py generate --model Kolibri-1-MLX-4bit --prompt "Explain briefly the difference between weather and climate." --chat-template-config '{"reasoning_effort": "low"}' --temp 1.0 --top-p 0.97 --top-k 128 --max-tokens 1024

For companies: your own GPU server (official route; not tested by us)

Aleph Alpha documents aleph-alpha-inference, which installs its vLLM plugin and exposes an OpenAI-compatible API. On suitable GPU hardware, install and serve:

python -m pip install "aleph-alpha-inference>=1.0"

vllm serve Aleph-Alpha/Kolibri-1 --kv-cache-dtype fp8 --reasoning-parser kolibri1 --tool-call-parser kolibri1 --enable-auto-tool-choice

Its alternative container is ghcr.io/aleph-alpha/aleph-alpha-inference. Recommended sampling is temperature 1.0, top_p 0.97 and top_k 128. Native training context is 262,144 tokens; quality was validated up to 1,048,576, but Aleph Alpha recommends at most 262,144 for complex work and serving efficiency. A large window does not guarantee perfect recall and uses more memory.

FAQ

Is it free? As of 8 October 2026, published weights and configuration carry an Apache 2.0 licence without a licence fee. Hardware, power and operations cost money; verify current hardware prices with the seller.

Does it speak Portuguese? Not as a declared target language. The card lists German and English. Fine-tuning is permitted by the licence but does not guarantee Portuguese quality.

Does it work in Ollama or LM Studio? Do not assume it does. As of 8 October 2026 the documented routes are vLLM with Aleph Alpha's plugin or community MLX with its launcher. We have not validated a one-click install in either app.

Can I put it in my product? Yes, Apache 2.0 weights permit commercial use, subject to licence conditions and responsible-use guidance. Data stays in-house only if your actual deployment does not forward prompts to a third party.

Has Cohere acquired Aleph Alpha? The companies announced an agreement on 16 September 2026. Regulatory approval was still pending as of 8 October 2026; do not treat the merger as completed.

Sources: official model card, community conversion, launch and vLLM commands and Cohere agreement.