What is Kolibri 1?
Kolibri 1 is an open-weight reasoning model from Aleph Alpha, the Heidelberg lab, released on 3 October 2026. Aleph Alpha is being absorbed by Canada's Cohere: the two signed a merger agreement on 16 September 2026, the combined company will trade as Cohere with headquarters in Berlin and Toronto, and Heidelberg stays as a research centre. The deal was still awaiting regulatory approval when we checked.
Kolibri is a mixture-of-experts model with 78.1 billion parameters, of which 3.46 billion are active per token, built for German and English only. It was pre-trained on 20 trillion tokens (about 62.5% English, 23.9% German and 13.6% code) on 768 NVIDIA B200 GPUs in Germany and Finland, with a tokeniser designed for German word structure. Its native context is 262,144 tokens, and Aleph Alpha says it has validated quality up to 1,048,576. It offers reasoning effort from none to high and Hermes-style tool calling, with a knowledge cut-off of 18 June 2026.
The FP8 weights are on Hugging Face under Apache 2.0 and take about 78 GB, so the minimum is one H200 or B200, or two 80 GB A100s or H100s. Aleph Alpha is a signatory of the EU's general-purpose AI Code of Practice. There is no public pay-per-token API.
Is Kolibri 1 worth it?
Kolibri earns its place on provenance first and scores second. It is an Apache 2.0 model trained on European infrastructure under German law, with documented training data, energy (an estimated 950 MWh) and Code of Practice commitments, which is what public-sector and regulated buyers in Germany are asked to show. On capability it is competitive with the open mixture-of-experts models of its class rather than ahead of the open field.
Aleph Alpha's own tables put it at an overall 75.5 in English and 70.8 in German, a shade above Qwen3.5 35B-A3B (74.7 and 69.8) and GPT-OSS 120B (72.3 and 70.2). The same tables show the dense Qwen3.8 27B at 80.2 and 79.9, so a German-first buyer gets better German from a Chinese model that is a third of the size, at the price of more compute per token. Every figure is self-reported with Kolibri at high reasoning effort, and no Artificial Analysis entry existed at launch.
The weak spots are agents and recall. SWE-Bench Verified is 66.4 against 73.8 for Qwen3.6 35B-A3B, TerminalBench 2.1 is 27.7 against 39.7 for Qwen3.5 35B-A3B, and on Aleph Alpha's German public-sector retrieval test it scores 75.0 to Qwen3.5's 80.0. It does better on not inventing answers: a 44.0 non-hallucination rate on the public AA-Omniscience set, against 11.1 for Qwen3.5 35B-A3B.
The engineering choices favour long documents. Only 10 of its 50 layers attend across the whole context, and Aleph Alpha says that on two H100s it serves 18 concurrent 256K-token requests where a 123B version managed 3. Its tokeniser packs 4.90 bytes of German per token against 4.35 for GPT-5's, so the same German text costs fewer tokens. Those gains assume you can run a plugin-patched vLLM on a 141 GB GPU or a pair of 80 GB ones.
The Cohere merger is the open question for buyers who chose Aleph Alpha for sovereignty. The Apache 2.0 weights already published cannot be withdrawn, but future Kolibri releases and support will come from Cohere, a Canadian company that will run dual headquarters in Toronto and Berlin.
Where it earns its place
An IT lead at a German federal agency
Run an internal assistant over agency documents without a US or Chinese model
Serve Kolibri on one H200 inside the agency's own data centre; the Apache 2.0 licence, European training and Code of Practice signature answer most procurement questions. Aleph Alpha's German public-sector retrieval score of 75.0 is respectable but trails Qwen3.5 35B-A3B's 80.0.
A platform engineer at a Munich insurer
Process 200-page German policy files in one request
Contexts up to 262,144 tokens are the recommended range, and the sliding-window design lets two H100s hold 18 such requests at once according to Aleph Alpha. Beyond that, add the --max-model-len 1048576 override.
A procurement manager comparing sovereign AI vendors
Judge whether committing to Kolibri now is safe given the Cohere deal
Weights already released under Apache 2.0 stay usable indefinitely, so self-hosting carries little lock-in. Support, future versions and any hosted offer will come from the merged Cohere once regulators approve, so contracts should name the successor entity.
Before you commit
- Setup
- pip install aleph-alpha-inference (or pull ghcr.io/aleph-alpha/aleph-alpha-inference), then vllm serve Aleph-Alpha/Kolibri-1 with an FP8 KV cache and the kolibri1 reasoning and tool-call parsers; plain vLLM cannot load it.
- Integrations
- OpenAI-compatible chat completions through vLLM, Hermes-style tool calling, and reasoning_effort set to none, low, medium or high in the chat template; Transformers also loads it. Aleph Alpha publishes its eval-framework for reproducing the benchmarks.
- Data and privacy
- Apache 2.0 with no separate acceptable-use policy and no revenue cap. Self-hosted, prompts stay on your servers; Aleph Alpha says customer-specific work never trains on customer data.
- Cost at scale
- The FP8 weights take about 78 GB. Aleph Alpha lists one H200, B200 or B300, or two 80 GB A100s or H100s, as the minimum, and recommends two H100s or two H200s for serving. No hosted per-token price is published.
- Limits
- German and English only; knowledge cut-off 18 June 2026; 262,144-token native context with 1,048,576 validated; recommended sampling is temperature 1.0, top_p 0.97, top_k 128. No independent leaderboard results at launch.
What's good
- Apache 2.0 with documented European training and Code of Practice commitments
- Only 3.46B active parameters per token; 1M-token context validated
- German tokeniser at 4.90 bytes per token, ahead of GPT-5 and Qwen
- AA-Omniscience non-hallucination rate of 44.0 (self-reported)
What's not
- German and English only
- Behind Qwen3.8 27B on Aleph Alpha's own German and English averages
- Needs about 78 GB of GPU memory and a custom vLLM plugin
- Pending Cohere merger clouds the long-term roadmap
Best for
German public bodies and regulated firms that need an auditable, European-trained open model for document and retrieval work on their own GPUs.
Kolibri 1 alternatives
Kolibri 1: common questions
- Is Kolibri 1 worth it?
- Kolibri is the strongest case yet for a European-built open model, though the best open model on Aleph Alpha's own German tests comes from Alibaba. We rate it 3.9 out of 5.
- How much does Kolibri 1 cost?
- Free under Apache 2.0. No public hosted API; enterprise deployment through Aleph Alpha sales, prices not published.
- What are the downsides of Kolibri 1?
- German and English only; behind Qwen3.8 27B on Aleph Alpha's own German and English averages; needs about 78 GB of GPU memory and a custom vLLM plugin; pending Cohere merger clouds the long-term roadmap.
- Who is Kolibri 1 best for?
- German public bodies and regulated firms that need an auditable, European-trained open model for document and retrieval work on their own GPUs.
- How much does Kolibri 1 cost for a team or in production?
- The FP8 weights take about 78 GB. Aleph Alpha lists one H200, B200 or B300, or two 80 GB A100s or H100s, as the minimum, and recommends two H100s or two H200s for serving. No hosted per-token price is published.
- Is Kolibri 1 safe for company data?
- Apache 2.0 with no separate acceptable-use policy and no revenue cap. Self-hosted, prompts stay on your servers; Aleph Alpha says customer-specific work never trains on customer data.
- What are the best Kolibri 1 alternatives?
- The closest alternatives we have reviewed are gpt-oss-120b, Qwen3.8-27B, Mistral Small 4, Command A+ and Mistral Medium 3.5.
Answers checked Oct 28, 2026.