Kolibri open weight: Aleph Alpha ships a 78.1B MoE under Apache 2.0

Kolibri open weight is Aleph Alpha’s October 3, 2026 drop: a sovereign-flavored English–German Mixture-of-Experts model with 78.1B total pa## Benchmarks without the cosplay

From Aleph Alpha’s published tables:

Overall averages on their harness read 75.5 (EN) and 70.8 (DE). Kolibri Origin sits lower — AIME 2026 at 81.5 and LiveCodeBench v6 at 59.2 — showing the jump from a 30.6B / 3.27B active precursor.

Kolibri open weight is not marketed as a Terminal-Bench assassin. 27.7 is an official number, not a crown. The pitch is Pareto efficiency: quality versus serving cost with only ~3.5B active parameters, plus German specialization and grounding work aimed at regulated customers.

How they claim they built it

The blog’s timeline is concrete. Pipeline work started in January. Kolibri Origin finished pre-training June 11, 2026. Kolibri finished September 11 after 20T pre-training tokens. Raw data processed is described as over 200T before filtering. German share is about 21.

Where this sits in the open-model pile

Readers tracking Memcool’s AI lane already bounced between Kimi K2.6, Gemini 4 Argon, and IBM agent identity. Kolibri open weight is less chatbot of the week and more on-prem MoE with a German specialization story.

Practical constraints apply. FP8 weights around the ~78 GB class and multi-GPU serving are not laptop toys. The serving examples assume vLLM plus Aleph Alpha’s inference plugin. If you need bilingual DE/EN with Apache 2.0 weights, Kolibri is aimed at you.

What to do as a reader

• Start from the official Aleph Alpha post and the HF repo.
• Keep Origin (30.6B / 3.27B) separate from Kolibri (78.1B / 3.46B).
• Note native 262k versus up-to-1M configuration flags.
• Evaluate TerminalBench 2.1 27.7 honestly.

Kolibri open weight is a real Apache 2.0 MoE drop with published architecture, a dated pretrain finish, and a bilingual specialization story. Download it, bench it on your tasks, and ignore anyone who turns a vendor table into destiny.3%.

Mid-training and long-context stages add more tokens. Knowledge cutoff is EN/DE: 18 Jun 2026. That process detail is part of the sovereign branding: training in Germany and Finland, transparent weights, and deployment freedom.rameters and about 3.46B active per token, published on Hugging Face under Apache 2.0. The company framed the release around German Unity Day, on-prem deployment, and regulated workloads.

• Official card: 78.1B total / ~3.46B active; context up to 1M (native training to 262k).
• Benchmarks: AIME 2026 96.0, GPQA Diamond 84.3, LiveCodeBench v6 85.9, HumanEval+ 92.7, SWE-Bench Verified 66.4, TerminalBench 2.1 27.7.
• Kolibri Origin: 30.6B total / 3.27B active; Kolibri pretrain finished September 11 on 20T tokens.

What actually happened

Aleph Alpha’s official blog states that Kolibri open weight ships full weights on Hugging Face with Apache 2.0 terms. The model is bilingual by design (DE/EN), with reasoning effort modes and tool calling via the aleph-alpha-inference stack and vLLM.

Architecture highlights: 50 MoE layers, 384 experts with 6 active plus a shared expert, and hybrid attention. Serving guidance points at FP8 KV cache and optional flags toward 1,048,576 tokens after native training through 262,144.

AICoder’s October 4 note mirrors the HF id Aleph-Alpha/Kolibri-1 and stresses native 262k context with extrapolation toward 1M.