Research · October 3, 2026 · 9 min read
On German Reunification Day, Aleph Alpha released Kolibri, a from-scratch 78-billion-parameter MoE with 3.46 billion active weights and a one-million-token context window under Apache 2.0, and its public benchmark tables put it ahead of Qwen, Mistral, and Nemotron models in the same active-parameter class on composite English and German scores.
On October 3, 2026, Aleph Alpha posted on X that “Kolibri is here”: seventy-eight billion parameters, three point four six billion active, up to one million tokens of context, “built in Europe,” with weights on Hugging Face under Apache 2.0. The same morning the lab published a blog post timed to German Reunification Day and a fifty-page tech report linked from the model card. For readers who follow open weights through Qwen, Mistral, and NVIDIA’s Nemotron lines, the drop can feel abrupt. Aleph Alpha has sold sovereign enterprise AI in Germany for years, but it has not been the default name in the MoE weight wars. Kolibri changes that by publishing a full checkpoint, an inference stack, and tables that compare the model directly to GLM, Nemotron, Qwen3.5 and Qwen3.6, Mistral Small, and Gemma variants at roughly three to six billion active parameters per token.
What shipped on Hugging Face
The release artifact is Aleph-Alpha/Kolibri-1 on Hugging Face, dated October 3, 2026. Kolibri is a mixture-of-experts transformer aimed at German and English, with explicit reasoning mode and Hermes-style tool calling. Total parameters are seventy-eight billion; active parameters per token are about three point four six billion, routed across three hundred eighty-four experts per layer with one shared expert and six routed. Weights ship in FP8 blocks with a documented memory footprint near seventy-eight gigabytes, and Aleph Alpha recommends hardware such as two H100s, one H200, or one B200 class GPU for serving.
Context is part of the headline. The model card validates quality up to one million tokens and recommends two hundred sixty-two thousand tokens or less for latency-sensitive or complex production work. Aleph Alpha pre-trained on sixteen-thousand-token sequences, mid-trained on sixty-five thousand, and extended to two hundred sixty-two thousand in a long-context phase, then reports extension without position scaling in sliding-window layers. The X thread’s one-million-token line matches the card, not a rounding error.
Trained from scratch, not repackaged
Aleph Alpha stresses a full training stack rather than a distill-and-release maneuver. The Hugging Face card says pre-training used twenty trillion tokens from a filtered bilingual corpus, roughly sixty-two point five percent English, twenty-three point nine percent German, and thirteen point six percent code, plus mid-training on three point four four trillion tokens and a long-context extension on two hundred one billion tokens. Post-training combined supervised fine-tuning on bilingual open and synthetic data with reinforcement learning across reasoning, agentic, and instruction environments. Compute disclosure on the card lists seven hundred sixty-eight NVIDIA B200 GPUs, twenty-one days of pre-training in measured wall time, and additional mid-training and long-context passes on the same cluster class.
The company’s blog describes Kolibri Origin, a thirty-billion-parameter, three-billion-active predecessor with a sixty-five-thousand-token context, as the pipeline proof. Hundreds of ablation runs and automated recovery when hardware or data links failed, in Aleph Alpha’s telling, preceded the larger model. Kolibri Base appears in evaluation tables as the checkpoint post-training starts from, with Kolibri Origin Base as the earlier ablation line. That narrative is how Aleph Alpha answers the “where did this come from?” question: not an overnight download of someone else’s weights, but a second-generation product from a training system the lab had already exercised in public on a smaller bird.
MoE benchmarks against Qwen, Mistral, and Nemotron
The comparative claim lives in Aleph Alpha’s own evaluation appendix on Hugging Face, with a footnote that MoE models are marked against each other while dense models are greyed out because they activate far more parameters per token. Within the three-to-six-billion-active MoE group, Kolibri reports an overall English composite of 75.5 and overall German of 70.8 after post-training. The same table gives Qwen3.5 35B-A3B 74.7 English and 69.8 German, Nemotron 3 Nano 30B-A3B 65.6 and 59.3, and Mistral Small 4 119B-A6B 63.1 and 61.4. Qwen3.6 35B-A3B sits between Kolibri and Qwen3.5 on those aggregates at 71.4 English and 67.3 German. Nemotron 3 Super 120B-A12B, with roughly twelve billion active parameters, scores 73.0 English and 67.9 German, still below Kolibri on those unweighted category means.
Task-level rows show the same pattern in places and gaps elsewhere. Kolibri posts 96.9 on AIME 2025 English against Nemotron Nano 89.6 and Mistral 79.8, and 84.3 on GPQA Diamond English against Qwen3.5 83.8. On SWE-Bench Verified it reports 66.4 versus Qwen3.5 71.6 and Nemotron Nano 38.6, so coding-agent leaderboard bragging rights are not uniform. TerminalBench 2.1 is 27.7 for Kolibri against 39.7 for Qwen3.5. Dense Qwen3.8 27B, outside the MoE comparison set, still shows higher overall composites in the same document at 80.2 English and 79.9 German. Afterimage is reporting Aleph Alpha’s table, not rerunning it. The fair summary is that Kolibri leads the MoE pack on the lab’s chosen averages while stronger dense baselines and individual agent benchmarks remain contested.
Aleph Alpha’s blog frames the same numbers as a Pareto frontier on quality versus decoding throughput per GPU for English and German, arguing no compared post-trained model delivers more benchmark score at the same serving cost. That chart is qualitative packaging on top of the Hugging Face tables. The tech report at aleph-alpha.com/downloads/tech-report.pdf carries architecture detail: fifty-layer MoE, four-to-one sliding-window to global attention, Muon optimizer, and Exact Quantile Balancing across experts.
Sovereign positioning and who gets to run it
Marketing language on Aleph Alpha’s site ties Kolibri to regulated German and European customers: public administration, industrials, aerospace, on-prem deployment, and supply-chain documentation from data ingestion through evals. That is a different pitch from a Silicon Valley API launch, even though the license is permissive. Apache 2.0 means weights are yours to host; the lab also ships an aleph-alpha-inference vLLM plugin and container image for reasoning and tool parsers.
Signatory status on the EU GPAI Code of Practice appears on the model card. Energy estimate on the card is about nine hundred fifty megawatt-hours for pre-training, mid-training, and long-context training combined, excluding SFT and RL. Those disclosures read like an enterprise vendor answering procurement questionnaires, which fits Aleph Alpha’s historical buyer more than a hobbyist download crowd.
Why the release feels like it came from nowhere
Until October 3, the open-weight conversation in late 2026 was still largely Alibaba, Mistral, NVIDIA, and Meta-shaped. Aleph Alpha was known for Pharia and government-facing rhetoric, not for topping Hugging Face trending with a seventy-eight-billion-parameter MoE. Kolibri lands with weights, code to serve them, a tech report, and tables that explicitly line up against Qwen3.5, Qwen3.6, Nemotron Nano and Super, and Mistral Small. The lab did not announce a closed preview months ahead; it paired a holiday blog post with immediate download.
Whether that counts as state of the art depends on which column you sort. On Aleph Alpha’s MoE-only composites in English and German, Kolibri is at the top of the named rivals with similar active width. On agentic coding tasks in the same document, Qwen still wins some rows. On dense models with more active parameters per token, Qwen3.8 27B remains ahead on overall averages. What is new is the combination: European lab, from-scratch training disclosure, Apache weights, one-million-token ceiling, and benchmark pages that read like a deliberate challenge to the usual open-weight leaderboard. The hummingbird logo on the card is small; the parameter table underneath is not.