Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsKolibri is an English-German language model from Germany’s Aleph Alpha, announced on 3 October 2026. Its weights are downloadable under Apache 2.0 terms. It is a 78-billion-parameter Mixture-of-Experts model that uses only 3.46 billion parameters for each token. It is built for regulated, on-premises work such as public administration, industry and aerospace. It is not a general “US model killer”: the benchmark evidence so far comes from Aleph Alpha itself and is mixed once you look past the headline math scores.
What Kolibri is, in numbers
Aleph Alpha describes Kolibri as an English-German Mixture-of-Experts (MoE) Transformer. Every figure below is the company’s own, taken from its launch article, product page and Hugging Face model card.
As an Amazon Associate I earn from qualifying purchases.
| Attribute | Value (as stated by Aleph Alpha) |
|---|---|
| Announced | 3 October 2026 |
| Languages | English and German |
| Total / active parameters | 78B total, 3.46B active per token |
| License | Apache 2.0, with full weights downloadable |
| Maximum context | 1,048,576 tokens |
| Recommended context | Up to 262,144 tokens for serving efficiency and complex tasks |
| FP8 weight footprint | About 78 GB |
| Pre-training | 20T tokens: about 62.5% English, 23.9% German, 13.6% code |
| Knowledge cutoff | 18 June 2026, for both languages |
“Open-weight,” not necessarily “open-source”
You can download and run the weights, and the license is permissive. The published material does not say the full training corpus is available. The accurate label is therefore open-weight. Anyone who needs to audit exactly what the model learned from cannot do so with the information released.
Why 78B total but 3.46B active matters
An MoE model routes each token through a small subset of its parameters. That lowers the computation per token, which is the basis of Aleph Alpha’s efficiency pitch. It does not lower memory needs, because every expert has to be loaded. Hence the roughly 78 GB for FP8 weights alone. Fast generation and small memory are different things.
#1 Best Overall
What Kolibri is designed to do
Aleph Alpha names these target uses:
- German- and English-language reasoning
- Coding
- Document processing and structured extraction
- Retrieval-augmented generation (RAG)
- Tool-using and agentic workflows
The model card specifies that it is intended for systems where a person reviews outputs before action. It is not meant for autonomous systems that act unreviewed. The company positions it for regulated and mission-critical settings and says customers can deploy it on premises. In its words: “Kolibri is a specialized language model built for sovereign mission-critical work in regulated areas including public administration, industrials and aerospace.”
Context window: one million tokens versus 262,144
The two context figures describe different things. One million tokens is the stated maximum, and Aleph Alpha says it validated quality and serving efficiency up to that length. The recommended ceiling for efficient serving and complex tasks is 262,144 tokens. This matches the final long-context training phase.
The model card gives the training progression: pretraining at 16,384 tokens, mid-training at 65,536 and a last long-context phase at 262,144. Contexts beyond 262,144 need explicit configuration in the serving setup. For most document and RAG work, plan around the 262,144 figure and treat the full million as a special case to test on your own material.
Rank #2
What it takes to run it
This is enterprise-class hardware territory. Aleph Alpha’s minimum configurations are:
- Two NVIDIA A100 80 GB GPUs, or
- Two H100 SXM5 GPUs, or
- One H200, B200 or B300
Its recommended configurations are higher. A laptop or single consumer GPU is not a realistic target. The 78 GB of FP8 weights alone fill most of a single 80 GB card before any context cache is counted, which is why the minimums are set where they are.
Serving uses Aleph Alpha’s own aleph-alpha-inference package with its vLLM plugin. The launch article provides a serving command. Check it there rather than relying on a copy, since the tooling is new and may change. Expect a specialized setup, not a one-line install of a stock runtime.
How good is it? Treat the benchmarks as claims
Aleph Alpha compares Kolibri with Qwen3.6 35B-A3B, Nemotron 3 Super 120B-A12B and Mistral Small 4 119B-A6B. It reports these headline results:
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →| Benchmark | Kolibri score (Aleph Alpha-reported) |
|---|---|
| AIME 2025 | 96.9 |
| AIME 2026 | 96.0 |
| AIME 2025, German | 87.5 |
| AIME 2026, German | 90.0 |
These are the publisher’s numbers, not independent results. No outside evaluation had been identified as of 5 October 2026, two days after release.
The published table also shows mixed outcomes. Kolibri leads on some measures, and rivals lead on others. Qwen3.6, for example, scores higher on several of the reported tool-use and knowledge comparisons. Nothing in the table supports calling Kolibri best overall. Aleph Alpha’s own conclusion is that Kolibri sits on a quality-versus-serving-cost Pareto frontier. It also says the model can match models with up to four times its active parameter count on selected math, coding, grounding and long-context tasks. Read that as a claim about selected tasks, based on the company’s own comparisons.
The named comparison models are all open-weight. The published comparison does not set Kolibri against the leading closed US systems. That is a gap to keep in mind when reading the “answer to the US” framing.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Kolibri against US models: how to compare fairly
Judging by European origin or by one score will mislead you. Compare on these axes instead:
- Language: German quality is the model’s clearest differentiator, and the German AIME results are reported alongside the English ones.
- Task: math, coding, grounding, long-context retrieval and tool use all rank the models differently.
- Serving cost and throughput: low active parameters help here, but the GPU memory bill remains.
- Context length: the practical 262,144 limit versus the validated one million.
- License: Apache 2.0 is permissive for commercial use and modification. Check each competitor’s terms rather than assuming parity.
- Data control: self-hosting keeps prompts and documents inside your own infrastructure. A hosted US API does not, whatever its contractual protections.
The data-control point is where the open-weight approach matters most. It is a deployment property, not a quality score. It applies to any model you can run yourself, though Kolibri is notable for being a recent German-English one.
Best Value
What “sovereign” does and does not mean here
“Sovereign” is Aleph Alpha’s positioning. Its stated rationale includes control over deployment and over the model supply chain. The published sources do not independently verify the company’s sovereignty, compliance or intellectual-property claims. A regulated buyer should still do its own review of licensing, training-data provenance and applicable rules. Apache 2.0 weights give you control over where the model runs. They do not by themselves give you a compliance certificate.
Training data and knowledge limits
The model card describes a filtered bilingual corpus of 20T tokens, plus mid-training and long-context phases. The split is about 62.5% English, 23.9% German and 13.6% code. These are model-card disclosures, not an audited dataset analysis. The built-in knowledge ends on 18 June 2026 for both languages. Anything newer must come from retrieval or tools, so a RAG or tool layer is needed for current facts.
Who should consider it
- A good fit: organizations with GPU infrastructure that need on-premises German-English document processing, extraction, RAG or reviewed agent workflows. These include public bodies and industrial firms.
- A weaker fit: individuals wanting a local assistant on modest hardware, teams that need many languages beyond English and German, and unsupervised autonomous agents, which the model card says are not the design target.
The practical next step is to run your own German and English documents through it, at the context lengths you will actually use. Compare the results against your current model before accepting any published ranking, Aleph Alpha’s included.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
