Aleph Alpha Releases Kolibri-1: 1M-Token Context, Open Weights—and Server-Grade Hardware
Aleph Alpha [released Kolibri-1 on October 3](https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/), publishing the weights under Apache 2.0. The English-German mixture-of-experts model has **78.1 billion parameters in total**, but activates **3.46 billion per token**. It supports explicit reasoning modes and tool calling, and the company says it validated a context length of up to **1,048,576 tokens**.
That combination makes the release relevant to teams exploring models they can run on infrastructure they control. It does **not** make Kolibri-1 a small laptop model. The active-parameter number describes how much of the network participates in a token's computation; it does not describe the memory needed to hold the full network.
What can developers use it for?
The [model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) names multi-step reasoning, retrieval over documents, agentic tool calling, and coding among intended uses. Aleph Alpha supplies an inference package with a vLLM plugin and examples for an OpenAI-compatible API, reasoning settings, and tool calls. A team could evaluate Kolibri for an internal code assistant, repository question answering, or a tool-using development workflow—provided it can operate the model and independently test its behavior on its own tasks.
The company also reports coding benchmark scores in its release post: **85.9 on LiveCodeBench v6** and **92.7 on HumanEval+**. These are *Aleph Alpha's published evaluations*. They are not tests we reproduced and do not measure whether a particular coding agent will apply a correct diff, use tools safely, or pass your repository's test suite.
The hardware catch behind “3.46B active”
Kolibri-1's [model card](https://huggingface.co/Aleph-Alpha/Kolibri-1) lists an approximate **78 GB FP8 weight footprint**. Minimum configurations include **two A100 80 GB GPUs, two H100 SXM5 GPUs, or one H200, B200 or B300**. The card explicitly explains the MoE trade-off: lower computation per token, while the **full model must be held in memory**. Those published minimums are server-grade, not a recommendation for a typical MacBook or PC.
The 1M-token figure also needs context. Aleph Alpha says the model was trained to 262,144 tokens in its final long-context phase and validated extension up to 1,048,576. For serving efficiency and complex tasks, the model card recommends **at most 262,144 tokens**. A long accepted prompt is not proof that sending an entire repository is the best way to solve a coding problem; large contexts still have memory, latency and retrieval trade-offs.
Is it free to use?
The weights are downloadable under Apache 2.0. The release and model card do **not** publish a verified hosted Kolibri-1 API price, so we cannot quote one. Self-hosting avoids a model-provider charge per generated token but comes with GPU, electricity and operational costs. Teams should also review the model card's license scope and any separate dependencies used to serve it.
**Bottom line:** Kolibri-1 is a noteworthy open-weight release for controlled, English-German deployments. Its coding capability merits a test on real development tasks, but “few active parameters” should not be mistaken for “runs on a laptop.” For a practical laptop-first workflow, start with a substantially smaller local coding model and measure the result on your own codebase.