Kolibri review

A free German model, 78B in size but only about 3.5B at work per word, built to run on company servers.

Out nowAleph AlphaFree AI model (download)Updated Oct 26, 2026
Download KolibriFree download on Hugging Face · official source, huggingface.co/Aleph-Alpha/Kolibri-1
Released3 Oct 2026
PriceFree (Apache 2.0)
Runs on1-2 server GPUs
First lookOur take, based on published facts and early independent tests

A free model from Germany, licensed for business use, that companies can run on their own servers. Excellent at math and German and cheap to run for its size, but weaker at long tool use and coding agents.

What is Kolibri?

Kolibri (Kolibri 1) is a free AI model from the German company Aleph Alpha, released on 3 October 2026. It works in English and German and is built for governments and regulated companies that must keep data on their own servers. It is 78 billion parameters in size, but only about 3.5 billion work on each word, which makes it cheap to run.

Key features

  • Free to download and use, including for business (Apache 2.0 license)
  • English and German, with about a fifth of its training in German
  • Top math scores: 96% on AIME 2026
  • Reads up to 1 million tokens, best kept to about 262,000
  • Built to say "I don't know" when the answer is not in its sources

Specs and price

Made byAleph Alpha (Germany)
Released3 Oct 2026
Size78B parameters, about 3.46B at work per word
LicenseApache 2.0: free, business use allowed
LanguagesEnglish and German
How much it can readUp to 1,048,576 tokens (trained up to 262,144)
Knows events up to18 June 2026
Computer neededAbout 78 GB of GPU memory, for example 2x A100 80GB or 1x H200
DownloadHugging Face, Aleph-Alpha/Kolibri-1
Hosted APINo price published yet

Test scores

TestScorePublished by
AIME 2026Math competition96% (Qwen3.6 35B: 91%)Aleph Alpha
GPQA DiamondHard science questions84.3% (Qwen3.6 35B: 83.4%)Aleph Alpha
SWE-bench VerifiedFixing real code bugs66.4Aleph Alpha
MMLU-ProGeneral knowledge80% (Qwen3.6 35B: 84.3%)Aleph Alpha
Tau3-Bench (banking)Customer service agent tasks38.1% (Qwen3.6 35B: 10.6%)Aleph Alpha

Scores published by Aleph Alpha are the maker's own numbers. We will add our own results in the hands-on review.

How to download and run Kolibri

  1. Download the model from Hugging Face: Aleph-Alpha/Kolibri-1.
  2. Run it with vLLM plus Aleph Alpha's aleph-alpha-inference plugin, from pip or as a container image.
  3. Use a server with about 78 GB of GPU memory, such as 2x A100 80GB or 1x H200. It will not run on a normal laptop.
  4. Keep the context at 262,144 tokens or less unless you have tested longer.

Who it is for

Good fit

  • European companies and public bodies that need data to stay in-house
  • Teams working in German
  • Math and reasoning-heavy work

Skip it if

  • Laptop users: it needs a GPU server
  • Coding agents and long tool use: Qwen3.6 does better

Pros and cons

Pros

  • Free, with a business-friendly license
  • Strong German and math
  • Cheap to run for its size
  • EU AI Act documentation published

Cons

  • Needs about 78 GB of GPU memory
  • Text only, no images
  • Weaker general knowledge and long tool use than Qwen3.6
  • Aleph Alpha is merging with Cohere, so the brand's future is unclear

How it compares

Against Qwen3.6 35B, a popular free model, Kolibri wins on math, hard science and banking-agent tests, and loses on general knowledge and tool calling (BFCL v4: 61.4% against 67.2%).

Hands-on review In testing

We are running Kolibri on real tasks. Results will be added to this page, at the same address.

Real tasksCoding, writing, research, agents
SpeedHow fast it answers in practice
Cost per taskWhat a typical job costs
VerdictWho should switch, who should wait

Related models