# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/show-hn-germanys-new-sovereign-ai-model-kolibri-49943034

# Kolibri is an open\-weight LLM from Aleph Alpha for German and English

356 points · 142 comments

[Full discussion](<https://news.ycombinator.com/item?id=49943034>)

[Read original](<https://tej.as/blog/aleph-alpha-kolibri>)

Category: [Models & Products](<https://hacksnap.live/?category=models-products>)

## Skept-o-meter & Hotness

Skept\-o\-meter: Low\. Estimated from 10 comments\.

10 comments for the summary\.

Peak rank: \#1

Time in Top 10: 23\.0 hours

Hacksnap ranks recent stories first, then orders each group by points\. Peak rank uses all retained observations\. Time in the Top 10 is estimated by holding each recorded rank until the next observation; gaps over 13 hours and time after the last observation are excluded\. Movement between observations is unknown\.

122 recorded rank observations from 2026\-10\-03T15:01:04\.472963\+00:00 to 2026\-10\-10T23:01:00\.95027\+00:00\.

Hotness — latest 122 recorded Hacksnap ranks:

2026\-10\-03T15:01:04\.472963\+00:00: rank \#11

2026\-10\-03T16:01:56\.56175\+00:00: rank \#7

2026\-10\-03T17:01:36\.651036\+00:00: rank \#4

2026\-10\-03T18:01:58\.24207\+00:00: rank \#2

2026\-10\-03T19:01:19\.196589\+00:00: rank \#2

2026\-10\-03T20:00:57\.313759\+00:00: rank \#1

2026\-10\-03T21:01:04\.44439\+00:00: rank \#2

2026\-10\-03T21:20:01\.585706\+00:00: rank \#2

2026\-10\-03T22:00:15\.503126\+00:00: rank \#2

2026\-10\-03T23:00:55\.584832\+00:00: rank \#2

2026\-10\-04T08:01:00\.327459\+00:00: rank \#2

2026\-10\-04T09:00:59\.321023\+00:00: rank \#2

2026\-10\-04T10:00:42\.975336\+00:00: rank \#2

2026\-10\-04T11:01:03\.76577\+00:00: rank \#2

2026\-10\-04T12:01:45\.307985\+00:00: rank \#2

2026\-10\-04T13:00:49\.176267\+00:00: rank \#2

2026\-10\-04T14:00:56\.952543\+00:00: rank \#2

2026\-10\-04T15:00:54\.739637\+00:00: rank \#71

2026\-10\-04T16:00:51\.403054\+00:00: rank \#72

2026\-10\-04T17:00:43\.166713\+00:00: rank \#72

2026\-10\-04T18:00:46\.315535\+00:00: rank \#71

2026\-10\-04T19:00:54\.768841\+00:00: rank \#71

2026\-10\-04T20:00:51\.070442\+00:00: rank \#72

2026\-10\-04T21:00:28\.550532\+00:00: rank \#71

2026\-10\-04T22:01:36\.217651\+00:00: rank \#73

2026\-10\-04T23:01:09\.400767\+00:00: rank \#73

2026\-10\-05T08:01:51\.387003\+00:00: rank \#72

2026\-10\-05T09:01:40\.528475\+00:00: rank \#71

2026\-10\-05T10:00:56\.290772\+00:00: rank \#71

2026\-10\-05T11:01:05\.919518\+00:00: rank \#71

2026\-10\-05T12:00:56\.590881\+00:00: rank \#70

2026\-10\-05T13:01:05\.197156\+00:00: rank \#72

2026\-10\-05T14:01:04\.085718\+00:00: rank \#72

2026\-10\-05T15:00:50\.814571\+00:00: rank \#72

2026\-10\-05T16:02:09\.54139\+00:00: rank \#71

2026\-10\-05T17:00:35\.336076\+00:00: rank \#71

2026\-10\-05T18:00:20\.445491\+00:00: rank \#71

2026\-10\-05T19:02:43\.604975\+00:00: rank \#72

2026\-10\-05T20:01:09\.777459\+00:00: rank \#72

2026\-10\-05T21:01:47\.650671\+00:00: rank \#73

2026\-10\-05T22:02:59\.93099\+00:00: rank \#74

2026\-10\-06T08:02:11\.210914\+00:00: rank \#74

2026\-10\-06T08:02:14\.354697\+00:00: rank \#74

2026\-10\-06T09:01:21\.688765\+00:00: rank \#76

2026\-10\-06T10:00:58\.047773\+00:00: rank \#76

2026\-10\-06T11:00:41\.689929\+00:00: rank \#77

2026\-10\-06T12:00:41\.7705\+00:00: rank \#78

2026\-10\-06T13:01:24\.509519\+00:00: rank \#78

2026\-10\-06T14:00:26\.950817\+00:00: rank \#78

2026\-10\-06T15:00:49\.066414\+00:00: rank \#79

2026\-10\-06T16:00:58\.166747\+00:00: rank \#81

2026\-10\-06T17:01:17\.009511\+00:00: rank \#81

2026\-10\-06T18:01:43\.791429\+00:00: rank \#82

2026\-10\-06T19:00:46\.048609\+00:00: rank \#82

2026\-10\-06T20:01:14\.687294\+00:00: rank \#82

2026\-10\-06T21:02:13\.414382\+00:00: rank \#82

2026\-10\-06T22:00:40\.926271\+00:00: rank \#81

2026\-10\-06T23:03:26\.867445\+00:00: rank \#83

2026\-10\-07T08:01:41\.339804\+00:00: rank \#84

2026\-10\-07T09:02:38\.026986\+00:00: rank \#85

2026\-10\-07T10:01:35\.069531\+00:00: rank \#85

2026\-10\-07T11:02:07\.770084\+00:00: rank \#85

2026\-10\-07T12:01:14\.996697\+00:00: rank \#84

2026\-10\-07T13:00:47\.45912\+00:00: rank \#83

2026\-10\-07T14:01:18\.307739\+00:00: rank \#83

2026\-10\-07T15:01:57\.541362\+00:00: rank \#85

2026\-10\-07T16:01:28\.474022\+00:00: rank \#85

2026\-10\-07T17:01:15\.036908\+00:00: rank \#86

2026\-10\-07T18:00:53\.300079\+00:00: rank \#85

2026\-10\-07T19:01:49\.456657\+00:00: rank \#86

2026\-10\-07T20:01:59\.032487\+00:00: rank \#86

2026\-10\-07T21:05:36\.655392\+00:00: rank \#88

2026\-10\-07T22:02:27\.105241\+00:00: rank \#89

2026\-10\-07T23:01:17\.452894\+00:00: rank \#89

2026\-10\-08T08:03:55\.083808\+00:00: rank \#89

2026\-10\-08T09:03:58\.68898\+00:00: rank \#87

2026\-10\-08T10:04:05\.364235\+00:00: rank \#88

2026\-10\-08T11:02:04\.410307\+00:00: rank \#87

2026\-10\-08T12:02:36\.39618\+00:00: rank \#87

2026\-10\-08T13:03:09\.206401\+00:00: rank \#87

2026\-10\-08T14:03:30\.113384\+00:00: rank \#88

2026\-10\-08T15:02:54\.444758\+00:00: rank \#86

2026\-10\-08T16:03:13\.067807\+00:00: rank \#85

2026\-10\-08T17:03:47\.92465\+00:00: rank \#84

2026\-10\-08T18:01:50\.360431\+00:00: rank \#84

2026\-10\-08T19:02:55\.719274\+00:00: rank \#85

2026\-10\-08T20:01:36\.359022\+00:00: rank \#85

2026\-10\-08T21:03:34\.000719\+00:00: rank \#85

2026\-10\-08T22:01:13\.459732\+00:00: rank \#84

2026\-10\-08T23:01:25\.318392\+00:00: rank \#83

2026\-10\-09T08:01:53\.314021\+00:00: rank \#84

2026\-10\-09T09:02:59\.93426\+00:00: rank \#85

2026\-10\-09T10:02:22\.652527\+00:00: rank \#84

2026\-10\-09T11:02:04\.807282\+00:00: rank \#84

2026\-10\-09T12:02:16\.769295\+00:00: rank \#84

2026\-10\-09T13:02:30\.35034\+00:00: rank \#86

2026\-10\-09T14:01:34\.215134\+00:00: rank \#85

2026\-10\-09T15:01:27\.564392\+00:00: rank \#84

2026\-10\-09T16:02:06\.561056\+00:00: rank \#85

2026\-10\-09T17:03:06\.673648\+00:00: rank \#87

2026\-10\-09T18:01:02\.851052\+00:00: rank \#87

2026\-10\-09T19:01:11\.250291\+00:00: rank \#87

2026\-10\-09T20:01:32\.91907\+00:00: rank \#87

2026\-10\-09T21:02:08\.376144\+00:00: rank \#89

2026\-10\-09T22:01:19\.253075\+00:00: rank \#91

2026\-10\-09T23:01:31\.990912\+00:00: rank \#91

2026\-10\-10T08:01:00\.593376\+00:00: rank \#90

2026\-10\-10T09:01:45\.527975\+00:00: rank \#92

2026\-10\-10T10:00:53\.799353\+00:00: rank \#92

2026\-10\-10T11:00:35\.65851\+00:00: rank \#91

2026\-10\-10T12:01:11\.169647\+00:00: rank \#91

2026\-10\-10T13:00:54\.008377\+00:00: rank \#90

2026\-10\-10T14:01:43\.208766\+00:00: rank \#92

2026\-10\-10T15:01:24\.825188\+00:00: rank \#92

2026\-10\-10T16:01:10\.213328\+00:00: rank \#91

2026\-10\-10T17:00:48\.84189\+00:00: rank \#90

2026\-10\-10T18:00:57\.55132\+00:00: rank \#90

2026\-10\-10T19:00:52\.379463\+00:00: rank \#89

2026\-10\-10T20:00:21\.993912\+00:00: rank \#89

2026\-10\-10T21:01:10\.657738\+00:00: rank \#90

2026\-10\-10T22:00:50\.433569\+00:00: rank \#89

2026\-10\-10T23:01:00\.95027\+00:00: rank \#90

Kolibri's German tokenizer and sovereign deployment are real differentiators, but the article's own benchmarks show weaker memory, tool calling and coding, leaving its practical niche narrow\.

## The brief

Aleph Alpha's Kolibri is an open\-weight mixture\-of\-experts LLM for German and English, released under Apache 2\.0 with 78 billion total parameters but only about 3\.46 billion active per token\. The article, based on Aleph Alpha's technical report and model card plus one tokenizer experiment, argues its German tokenizer, long\-context attention, German\-language reasoning and abstention training make it suited to sovereign, document\-heavy German deployments\. It also reports clear weaknesses: weak closed\-book recall, multi\-turn tool calling and coding, and 78 GB of weight memory\. The author recommends it for RAG over long German documents on controlled hardware, not as a general\-purpose or coding model\.

- Architecture: 50 layers, 384 experts plus one shared expert per layer; the router sends each token to 6 experts, yielding 78\.1B total and 3\.46B active parameters\.
- Tokenizer: 128,000\-token UniBPE vocabulary; the author's test on the German Basic Law found 15% fewer tokens than GPT\-5's o200k\_base and roughly tied English\.
- Context: 40 of 50 layers use 512\-token sliding\-window attention, every fifth layer full attention; native 262,144 tokens, validated to 1,048,576\.
- Training: about 24T tokens, over one\-fifth German, on 768 NVIDIA B200 GPUs; English data rephrased with Gemma 4, German with Mistral\-NeMo, Qwen3\-32B labels for quality filters\.
- Evaluation: Aleph Alpha reports leading German/English scores among \~3B\-active MoE peers, including AIME 2025 German 87\.5 vs 84\.4, but weaker closed\-book RAGBench 51\.0, multi\-turn BFCL 39\.8 vs 58\.2, Terminal\-Bench 27\.7 vs 39\.7\.
- Deployment: needs \~78 GB GPU memory (2×A100/H100 80GB, H200/B200/B300), Aleph Alpha's vLLM plugin supporting one vLLM version, no hosted provider at launch, and only German/English\.

## Discussion themes

Analyzed: 2026\-10\-03T18:00:59\.906084\+00:00

Analysis sample: Based on 23 of 23 usable stored comments. Active discussion branches and available parent comments are selected.

This sample may omit parts of the full thread. Selected themes do not measure community opinion or how common a view is.

### Evaluation baseline fairness

A commenter criticizes the evaluation for omitting Qwen3\.8 Flash, another small\-active MoE, and for comparing Kolibri with the older Qwen3\-Next 80B\-A3B; they say this does not invalidate the model's point but weakens the comparison\.

Sources: [Comment 49944340](<https://news.ycombinator.com/item?id=49944340>)

### Sovereign AI as an auditing layer

A commenter proposes that a sovereign model's main near\-term use is auditing other models' tool calls and end results as a trust adapter for government applications, provided it is cheap and fast enough\.

Sources: [Comment 49944391](<https://news.ycombinator.com/item?id=49944391>)

### Sovereign AI market adoption and regulation

Comments debate whether European buyers will choose sovereign models without legal mandates, whether data\-location requirements drive enterprise demand, and whether long\-term sovereign investment can overcome US and Chinese momentum\.

Sources: [Comment 49944417](<https://news.ycombinator.com/item?id=49944417>) · [Comment 49944457](<https://news.ycombinator.com/item?id=49944457>) · [Comment 49944501](<https://news.ycombinator.com/item?id=49944501>) · [Comment 49944138](<https://news.ycombinator.com/item?id=49944138>) · [Comment 49944192](<https://news.ycombinator.com/item?id=49944192>)

### Open\-source ethical risks

A commenter warns that open\-source LLMs democratize access but also enable harmful actions at arbitrary scale, arguing that society lacks the ethical maturity to handle that power\.

Sources: [Comment 49945632](<https://news.ycombinator.com/item?id=49945632>)

### Token efficiency and inference speed

A user on an RTX Pro 6000 setup reports that the model spends too many tokens overthinking even when it finds the correct approach, while noting good speed of about 170 tokens per second in fp8 with 3B active parameters\.

Sources: [Comment 49943869](<https://news.ycombinator.com/item?id=49943869>)

### Technical report openness and reproducibility

A commenter praises the technical report for explaining the model and dataset in tutorial\-like detail, calling it an unusually high level of openness\.

Sources: [Comment 49944996](<https://news.ycombinator.com/item?id=49944996>)

## Sources & coverage

AI-generated summary · 2026\-10\-03T15:01:02\.861631\+00:00

Based on 10 of 10 usable stored comments, selected by depth and branch activity. This is a sample of the discussion. Article text may also be shortened.

Generated using deepseek\-ai/DeepSeek\-V4\.1\-Flash. Check the linked sources for full context.