Runs on ≤96 GB · Hugging Face ↗
Not independently scored by Artificial Analysis — its base model Aleph-Alpha/Kolibri-1-BF16 is the closest proxy
Aleph Alpha's reasoning flagship (Apache-2.0), and the first European model in this table at this size: 78.1B parameters of which 3.46B are active per token. A mixture of 384 experts with top-6 routing plus one shared expert, over 50 layers: 40 of them sliding-window attention with a 513-token window and 10 full attention. 262k native context; Aleph Alpha says it extends to 1M and recommends staying at 262k. Thinking has three effort levels, and reasoning_effort none (or enable_thinking=false) turns it off. The weights are published in full, FP8 for Kolibri-1 and BF16 for Kolibri-1-BF16, but the repository ships no modeling code and the architecture is in neither transformers nor upstream vLLM. Aleph Alpha serves it through its own vLLM plugin, the aleph-alpha-inference package. For a GGUF we converted and quantised the GGUF ourselves with a community llama.cpp patch (four commits, not merged upstream) that adds the kolibri1 architecture and its sigmoid-with-bias expert gating.
Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈1 GB at 32K, fp16). Read from the GGUF header, but this architecture has never been checked against llama.cpp — treat it as an estimate.
| Quantization | Size |
|---|---|
| Q2_K | 28.6 GB |
| Q3_K_M | 37.5 GB |
| Q4_K_M | 47.5 GB |
| Q5_K_M | 55.6 GB |
| Q6_K | 64.2 GB |
| Q8_0 | 83.1 GB |
Read HellaSwag 65.21 and WinoGrande 61.09 as 3B-active scores, because that is what those tasks measure, not as what 78B of weights can do. The 83.1 HellaSwag on the model card belongs to Kolibri Base, which Aleph Alpha has not published, so it is not a contrast with our number: the post-trained model's own table publishes no HellaSwag and no WinoGrande at all. Measured at Q4_K_M (44.2 GB) with 14 of 50 layers on a 24 GB card and the rest in system RAM. Because the patch is unmerged we checked the conversion before trusting the scores: wikitext-2 perplexity 19.67 +/- 0.17 over 144 chunks and coherent German and English answers, so the numbers are the model's, not a broken conversion's.