Aleph Alpha released Kolibri 1, a German-English open-weight language model with 78.1 billion parameters, on 3 October. Its Apache 2.0 licence allows users to modify and redistribute the weights, but running the full model still calls for substantial hardware.
Watch Desk analysis
What happened
The model activates 3.46 billion parameters per token and has a native context length of 262,144 tokens. Aleph Alpha recommends staying within that limit for latency-sensitive deployments and complex tasks. The FP8 weights take about 78 GB; the company’s listed minimum configurations include two 80 GB A100s or H100s, or a single H200, B200 or B300.
The architecture uses 50 layers, each with 384 routed experts and one shared expert. Only 10 layers read the full context; the other 40 use a 512-token window. ActuIA says the model card and technical report describe those design choices, while the company’s benchmark comparisons use its own evaluation harnesses.
Key findings
- Weights under Apache 2.0
Users can reproduce, modify and redistribute the weights, subject to the licence’s notice requirements. - Open weights are not the whole stack
The licence covers weights and configuration files, not the underlying code, architecture, parameter settings or training methods. - A sizeable memory footprint
The FP8 weights take about 78 GB, before accounting for the rest of a deployment. - German and English are the declared languages
Teams working with other languages, including French, will need to test suitability rather than assume coverage. - Vendor benchmarks have uneven results
ActuIA notes Kolibri’s results vary by task, including weaker scores than some comparisons on SWE-Bench Verified and factual error correction.
Why it matters
Kolibri gives organisations a model they can download and run under a permissive licence, rather than relying only on an API. That can matter for control over deployment and continuity, but the hardware requirements make “open” rather different from “easy to run”.
The release also arrives as Aleph Alpha is due to combine with Cohere, subject to regulatory approvals. The announced agreement says the combined company will operate under the Cohere name. That makes the licence boundary especially useful to understand: rights to this release do not promise future versions, maintenance or support.
Our read
Kolibri is a substantial option for teams whose work is mainly in German or English and who have the hardware to host it. The Apache licence is meaningful, but it does not turn the entire technology stack into open source or guarantee what happens next. Check the model card, test your workload and budget for the machine, not just the download.
What to watch
- Whether Aleph Alpha publishes further deployment guidance or independent evaluations.
- How Kolibri performs on real workloads beyond the company’s benchmark harnesses.
- Whether the Cohere combination closes and how the combined company handles future model releases.
Discussion spark: For organisations weighing local control against hosting costs, is a model with downloadable Apache-licensed weights worth the hardware burden when the rest of the stack is not licensed with them?
Sources and evidence
- Aleph Alpha's Kolibri: what the Apache 2.0 license guarantees – ActuIA (6 October 2026, 10:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.