Aleph Alpha open-sources Kolibri-1: a 78B MoE with only 3.46B active params
German AI company Aleph Alpha open-sourced its reasoning MoE model Kolibri-1 on Hugging Face on October 3 (German Unity Day), where it quickly trended. Under the Apache 2.0 license, it is positioned as a sovereign, European open-weight model: trained from scratch entirely in Germany, not a Qwen fine-tune, with a focus on German and English bilingual capability after months of pre- and post-training.
Confirmed
- Specs: 78.1B total parameters, only 3.46B active per token (4.4%), context length up to 1M tokens
- Architecture details (per @TejasKumar): 384 small experts per layer, with the router activating 6 per token — roughly 3.5B parameters running per word
- Benchmark performance (per @TejasKumar): 96.9% on AIME 2025, beating all tested MoE models of similar or larger size, including those with more active parameters per token
- Release format: available in vLLM and safetensors formats with an accompanying arXiv paper; weights can run on users' own hardware. Per @QuixiAI, the weights are in FP8 and run directly, and a feature request has already been filed for support in the open-source inference engine colibri (39.5k stars)
- The model repo includes chat template configuration supporting system prompts, reasoning effort settings, and tool-calling format (per @JiliJeanlouis)
- Third party Tesseracted Labs is hosting the model, letting users try it free in the browser for a few days without GPU or registration, with Brave Search web search enabled (per @FrankNoeBerlin)
Unconfirmed
- Benchmarks such as the 96.9% AIME score currently come mainly from the publisher's materials and relayers like @TejasKumar, and await independent third-party verification
Why it matters
- This is a flagship open-source result for Europe's "sovereign AI" route: trained and released entirely in Germany, offering Europe a controllable option independent of the US and China
- The very low activation ratio (4.4%) implies much lower inference cost; if the benchmarks hold up under third-party testing, it will be a strong contender in the efficient-reasoning track
- Free third-party trial access lowers the barrier to entry and helps the community build independent evaluations quickly
2026-10-03 ~ 2026-10-05 · 14 related posts
- Episode 1: Aleph Alpha open-sources Kolibri-1: a 78B MoE with only 3.46B active params(2026-10-03, 14 posts)
- Episode 2: Aleph Alpha Open-Sources Kolibri: 78B-Parameter Model with 1M Context(2026-10-05, 3 posts)
- Episode 3: Ex-AWS Evangelist Debunks Kolibri as Europe's DeepSeek Moment(2026-10-06, 2 posts)
Primary sources
- Aleph Alpha's sovereign open-weight Kolibri scores 96.9% on AIME, beating 3x bigger MoE models — TejasKumar_ · 2026-10-03
- Aleph Alpha launches Kolibri, an Apache 2.0 sovereign open-weight LLM built in Germany — TejasKumar_ · 2026-10-03
- [source] Aleph Alpha open-sources Kolibri-1: 78B MoE with 3.46B active params and 1M context — Nunki08 · 2026-10-03
- Aleph-Alpha's MoE reasoning model Kolibri-1 trends on Hugging Face — Aleph-Alpha · 2026-10-03
- [source] Aleph Alpha open-sources Kolibri: 78.1B MoE with 3.46B active params under Apache 2.0 — petrusenko_max · 2026-10-04
- Aleph Alpha open-sources Kolibri-1 model on Hugging Face — JiliJeanlouis · 2026-10-04
- Aleph Alpha Releases Kolibri: a From-Scratch 78B MoE With Only 3.46B Active Params, Apache 2.0 — solyarisoftware · 2026-10-04
- Aleph Alpha open-sources German sovereign model Kolibri-1, free to try in browser — FrankNoeBerlin · 2026-10-04
- Aleph Alpha open-sources Kolibri: 78B params, 3.46B active, 1M context, Apache 2.0 — josh_wills · 2026-10-04
- Aleph Alpha releases open-weight European model Kolibri-1 under Apache 2.0 — PMinervini · 2026-10-04
- [source] Aleph Alpha's Kolibri-1: 78B MoE with 3B active params, native FP8 weights — QuixiAI · 2026-10-05
- Stealthy German lab Aleph Alpha drops open-weight Kolibri: 78B params, 3.46B active, 1M context — skdh · 2026-10-05
2 near-duplicate retellings: TejasKumar_ · TejasKumar_