Kolibri Lands: A 78.1B Open-Weight Model Under Apache 2.0 That Activates Only 3.46B Parameters

On 3 October 2026 — Germany's Day of Reunification — Heidelberg lab Aleph Alpha released Kolibri: an English-German mixture-of-experts model with 78.1 billion total parameters, released as open weights on Hugging Face under the Apache 2.0 licence, no gate, no waitlist [1]. The number that matters is not 78.1 billion. It is 3.46 billion: the parameters that actually fire for each token, or 4.4% of the model [1][4]. That one design choice is why a model with a validated 1,048,576-token context window is quoted at roughly 78 GB of FP8 weights — small enough to think about hosting yourself.

What actually happened

The weights are public, and the licence is the permissive one. Kolibri-1 is downloadable from Hugging Face under Apache 2.0, alongside a BF16 checkpoint [4]. The licence grant is narrower than the headline suggests: it applies to the weights and configuration files in the repository, and the model card is explicit that it "does not extend to underlying code, model architecture, parameter settings or any training method." You can run it, fine-tune it and ship products on it; you cannot reproduce how it was built.

Only a slice of the model runs at any moment. Kolibri is a 50-layer transformer with 384 routed experts per layer, 1 shared and 6 routed, so each token activates 3.46B of 78.1B parameters [4]. Aleph Alpha's own benchmark tables, and the technical report behind them, put it on the Pareto frontier for quality versus serving cost in both English and German — meaning, in its words, that no compared model delivers more quality at the same serving cost or the same quality at lower cost [1][3]. On its reported numbers, Kolibri scores 96.9 on AIME 2025, 84.3 on GPQA Diamond, 85.9 on LiveCodeBench v6 and 66.4 on SWE-Bench Verified, while matching models with up to four times its active parameter count such as NVIDIA's Nemotron 3 Super [1]. In agentic tool use it posts 94.7 on τ²-bench telecom and 38.1 on τ³-bench banking, against 10.6 for Qwen3.6-35B-A3B [1]. These are vendor-published evaluations; independent third-party runs had not appeared in the days after release.

The context window is real, with a caveat attached. Native training reaches 262,144 tokens, and quality and serving efficiency have been validated up to 1,048,576 — but the model card recommends staying at or below 262,144 for latency-sensitive or complex work [1][4]. Beyond that length you pass --max-model-len 1048576 to vLLM with a positional-encoding override.

Serving it is a documented, ordinary workflow. The release ships an aleph-alpha-inference package that plugs into vLLM, exposed as an OpenAI-compatible endpoint, with reasoning effort selectable per request (none, low, medium, high) and tool calling enabled by the same command [1][4]. Hardware is the honest constraint: about 78 GB of FP8 weights means a minimum of 2× A100 80 GB, 2× H100 SXM5, or a single H200/B200/B300 [2][4].

The provenance is published, which is unusual. Pre-training used a 20T-token bilingual corpus (~62.5% English, ~23.9% German, ~13.6% code), plus 3.44T mid-training tokens and 201B for long-context extension [4]. The pre-training run itself used 768 NVIDIA B200s for 21 days — 392,000 GPU-hours — at an estimated 6.4e23 FLOPS, with total training energy put at roughly 950 MWh [4]. Aleph Alpha says the model was built in Germany and trained on infrastructure in Germany and Finland, designed with the EU AI Act, the General-Purpose AI Code of Practice and GDPR in mind, and that it is trained to abstain — to say it does not know — when the supplied context does not support an answer [1][4].

Why it matters for Indian SMBs and mid-market teams

• This is the first credible "own your inference" option at this capability tier. Apache 2.0 weights mean no per-token bill, no rate limits, no vendor deprecation notice, and no customer document leaving your network — the same argument that made Postgres and Redis default choices. For a firm handling contracts, filings or health records under India's DPDP Act, "the model runs inside our own VPC" is an answer to a question your auditor will eventually ask.
• Do the arithmetic before you get excited. Two A100 80 GB cards is a real machine, not a workstation. Renting that capacity full-time will beat an API bill for sustained, high-volume workloads and lose badly for a team that calls a model a few hundred times a day. The right question is not "is it free?" but "what is our tokens-per-day, and does self-hosting win at that volume?"
• The 4.4% active-parameter trick is the trend to watch. Sparse models that fire a fraction of their weights deliver large-model quality at small-model compute cost. Every serious lab now ships one. That shifts the build-versus-buy line: routine classification, extraction and drafting can move to a model you host, while the expensive frontier API is reserved for the hard ten percent.
• Licence hygiene still needs a human. "Open weights" is not "open source" here — the training code, architecture and methods are excluded [4]. If a vendor or consultant tells you the model is fully open, ask which artefacts the grant actually covers before you commit a product roadmap to it.
• Sovereignty is now a procurement feature, not a slogan. Aleph Alpha built Kolibri specifically for regulated public-administration and industrial buyers who cannot send data to a foreign inference service [1]. Indian mid-market firms selling into government, healthcare or finance will start seeing the same clause in tenders — having a worked-out private-deployment answer is becoming table stakes.

Where Guruji Corporation fits in

We evaluate every open-weight release against a client's actual workload rather than its benchmark table: tokens per day, latency budget, GPU availability, and the data-residency rule that governs the data itself. Kolibri is a strong candidate for teams that need English-and-local-language document work inside their own infrastructure — and a poor one for a team that needs a model on a laptop or a shared CPU server.

That assessment is what we do before any hardware is rented: a two-week pilot on your own documents, measured quality against your current provider, and a cost model showing the crossover volume where hosting wins. If self-hosting never wins for your workload, we tell you that too. Talk to us and we will run the numbers on your stack.

Sources - Kolibri Has Landed: A Sovereign Open-Weight Model — Aleph Alpha - Kolibri model page — Aleph Alpha - Kolibri: A Sovereign European Model on the Pareto Frontier (technical report) — Aleph Alpha - Aleph-Alpha/Kolibri-1 model card — Hugging Face