ModelsArchitectures & capability
Aleph Alpha releases Kolibri-1, a 78B-parameter German-English open-weight model under Apache 2.0
Heidelberg-based Aleph Alpha has published Kolibri-1, a German-English mixture-of-experts model with 78.1B total and 3.46B active parameters, under Apache 2.0. Its benchmark results against spring-2026 rivals are mixed, and running it requires data-centre GPUs.
QuietHow it’s measured
YouTube video · 133 views · 2 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
AT Pulse research deskPublished · 5 min read · AI-assisted analysisKolibri-1 is a text-only mixture-of-experts model for German and English. It has 78.1B total parameters, but only 3.46B are active for each token. Its native context is 262k tokens. It can reach about 1M tokens through extrapolation, but the model card advises staying at or below 262k. It offers four reasoning-effort levels and tool calling. Weights come in FP8 and BF16 versions. [7] [8]
All scores come from Aleph Alpha's own evaluation framework, run at the model's highest reasoning setting. Against Qwen3.6-35B-A3B, Kolibri leads on AIME and a banking tool-use test. It trails on MMLU-Pro and the BFCL v4 function-calling test. The rival models date from spring 2026, newer models are left out, and Artificial Analysis has not yet listed Kolibri. [4] [9] [12]
the weights are EU-trained, permissively licensed, can run offline and focus on German. Implication: regulated German-language buyers gain a sovereign option. They should still test it on their own workloads and weigh roadmap risk from the merger.
Executive brief
Aleph Alpha's new open-weight model, Kolibri-1, matches or beats several open rivals on Aleph Alpha's own benchmarks while using only 3.46B active parameters per token. No independent lab has verified those scores yet, and the main rivals in Aleph Alpha's comparison were released in spring 2026 (Startup Fortune). The Heidelberg lab released the model on 3 October 2026 under Apache 2.0. It has 78.1B total parameters and supports German and English. Aleph Alpha is also in the middle of a pending merger with Cohere. The Reddit post and video this dossier started from are promotional commentary that repeat vendor charts without testing the model.
What changed and event timeline
Merger with Cohere announced
Cohere said it would combine with Aleph Alpha, with Schwarz Group as the anchor financial backer ().
Merger agreement signed
The deal values the combined company at a reported ~$20B and still needs regulatory approval. The combined company would use the Cohere brand ().
Kolibri-1 released
The FP8 weights went up on Hugging Face under Apache 2.0, along with a technical report (;).
3–4 Oct 2026: First caveats from coverage
Trending Topics, as cited by others, pointed out that Kolibri has no Artificial Analysis listing and is compared against older rivals (;).
Commentary amplifies vendor charts
A Reddit post and an AI-avatar video framed the release as "beating Qwen, Nemotron and Mistral" (;).
Capabilities and access
- Versions:
Aleph-Alpha/Kolibri-1ships as FP8 weights, with aKolibri-1-BF16variant also listed (model card). - Size: 78.1B total parameters, 3.46B active per token.
- Context: 262,144 tokens natively. The 1,048,576-token window is reached by extrapolation, and the model card recommends staying at or below 262k.
Read the full section
- Versions:
Aleph-Alpha/Kolibri-1ships as FP8 weights, with aKolibri-1-BF16variant also listed (model card). - Size: 78.1B total parameters, 3.46B active per token.
- Context: 262,144 tokens natively. The 1,048,576-token window is reached by extrapolation, and the model card recommends staying at or below 262k.
- Features: Four reasoning-effort levels (none, low, medium, high), Hermes-style tool calling, text only, German and English. Knowledge cutoff is 18 June 2026.
- Hardware: The weights take about 78GB. The minimum is 2× H100/A100-80GB or one H200/B200, which means "run it at home in LM Studio" (video 05:38) is not realistic on typical consumer machines.
Technical analysis for researchers and developers
- Architecture: 50 MoE layers, each with 384 experts (6 routed per token plus 1 shared). There is a custom 128k tokenizer (model card).
- Training: 20T pre-training tokens, 3.44T mid-training and about 200B long-context tokens, on 768 B200 GPUs for about 21 days in Germany and Finland.
- Why 78B: Aleph Alpha says a 123B variant served only 3 concurrent 256k-token requests on two H100s, against 18 for the 78B model, which also decodes 28% faster.
Read the full section
- Architecture: 50 MoE layers, each with 384 experts (6 routed per token plus 1 shared). Attention mixes sliding-window and grouped-query layers at 4:1, with a 512-token window. There is a custom 128k tokenizer (model card).
- Training: 20T pre-training tokens, 3.44T mid-training and about 200B long-context tokens, on 768 B200 GPUs for about 21 days in Germany and Finland. Post-training used reinforcement learning (RL) on 1.2M tasks (Aleph Alpha blog; mer.vin).
- Why 78B: Aleph Alpha says a 123B variant served only 3 concurrent 256k-token requests on two H100s, against 18 for the 78B model, which also decodes 28% faster.
- Evaluation: Every comparison was run with Aleph Alpha's own evaluation framework, with Kolibri at its highest reasoning setting (Startup Fortune).
- Inconsistent German share: Sources disagree on how much of the training data is German: 23.9% (model card), about 21.3% (blog) or 23.4% (AI Weekly).
Claims and evidence
All of the scores below come from Aleph Alpha and have not been independently replicated.
Read the full section
All of the scores below come from Aleph Alpha and have not been independently replicated.
- Maths and science: AIME 2026 96.0, GPQA Diamond 84.3, HumanEval+ 92.7, SWE-Bench Verified 66.4 (model card).
- Head-to-head with Qwen3.6-35B-A3B (DataCamp):
- Kolibri leads on AIME and on Tau3-Bench Banking (38.1 vs 10.6).
- It trails on MMLU-Pro (80.0 vs 84.3), BFCL v4 (61.4 vs 67.2) and AA-Omniscience.
- Mixed tool-use results: Even on vendor numbers, the "tool chart" win in the video (03:25) is mixed.
- Fewer wrong answers: Kolibri abstains on 44% of AA-Omniscience items instead of answering incorrectly, compared with 15% for its predecessor (blog).
- No independent rating: Kolibri has no Artificial Analysis listing (Progressive Robot).
Context and prior work
- Predecessor: Kolibri-1 follows "Kolibri Origin," released 11 June 2026 (blog).
- Rivals: The comparison set is Qwen3.6-35B-A3B, Nemotron 3 Super 120B and Mistral Small 4 119B, all from spring 2026.
- Corporate future: The pending Cohere merger means the Aleph Alpha brand is set to be absorbed into Cohere (SiliconANGLE).
Read the full section
- Predecessor: Kolibri-1 follows "Kolibri Origin," released 11 June 2026 (blog).
- Rivals: The comparison set is Qwen3.6-35B-A3B, Nemotron 3 Super 120B and Mistral Small 4 119B, all from spring 2026. Newer models such as GLM-5.3 and Kimi K3 are left out (Progressive Robot).
- Corporate future: The pending Cohere merger means the Aleph Alpha brand is set to be absorbed into Cohere (SiliconANGLE). Startup Fortune describes the integration as already done, which conflicts with SiliconANGLE's report that regulatory approval is still pending.
Limitations, safety and contested findings
- Long context: RULER scores fall from 86.9 at 4K tokens to 63.2 at 1M tokens (Progressive Robot).
- Possible contamination: DataCamp reports 48% (English) and 47% (German) HumanEval contamination in the training data, which would inflate coding scores (DataCamp).
- Uneven performance by sector: On customer proxy benchmarks, Kolibri scores 59% in aerospace against 99% in automotive (blog).
Read the full section
- Long context: RULER scores fall from 86.9 at 4K tokens to 63.2 at 1M tokens (Progressive Robot).
- Possible contamination: DataCamp reports 48% (English) and 47% (German) HumanEval contamination in the training data, which would inflate coding scores (DataCamp). The reviewed sources did not independently confirm these figures.
- Uneven performance by sector: On customer proxy benchmarks, Kolibri scores 59% in aerospace against 99% in automotive (blog).
- Stated risks: The model card lists bias, outdated knowledge and anthropomorphization, and recommends human review for high-stakes use (model card).
- Video's own basis: The presenter says they have not tested Kolibri (05:00) and took the technical details from an AI summary of the report (07:38).
Business and practitioner implications
- Who it suits: German-language public-sector and regulated buyers get an Apache 2.0 model, trained in the EU, that they can run offline.
- Cost: Budget for data-centre GPUs. Serving efficiency at long context is the vendor's main economic claim.
- Before adopting:
Read the full section
- Who it suits: German-language public-sector and regulated buyers get an Apache 2.0 model, trained in the EU, that they can run offline. Aleph Alpha is an EU GPAI Code of Practice signatory (model card).
- Cost: Budget for data-centre GPUs. Serving efficiency at long context is the vendor's main economic claim.
- Before adopting:
- Benchmark it against current Qwen releases on your own German and tool-use workloads.
- Keep contexts at or below 262k tokens.
- Factor in the vendor roadmap risk from the Cohere merger.
The source trail.
Sources (13)
Aleph Alpha AI: Europe's New Free Rival to Qwen and Nvidia
Feed excerpt only. A transcript of the video linked from this page is supplied (linked video auto captions; https://www.youtube.com/watch?v=LyrreBHx6DA); automatic text may contain errors.
reddit.comWhy My Next Local AI Model Download Is a Free German Model
Related coverage; assess separately
reddit.comLinked video: Aleph Alpha AI: Europe's New Free Rival to Qwen and Nvidia
linked video auto captions
www.youtube.com