AI Tools

Sovereign AI in 2026: which open-weight models can you actually download?

Published 2026-10-13 · Researched and price-checked on official sources

Sovereign AI got its first downloadable flagship on 3 October 2026, when Aleph Alpha published Kolibri, a 78B Mixture-of-Experts model under Apache 2.0. Serving it needs at least two A100 80GB GPUs. Reflection's 501B Beam followed two days later, but its weights are still waitlist-only.

Key facts

What matters

Kolibri 1 (Aleph Alpha) was released on 3 October 2026: 78B total parameters, 3.46B active per token, a 1,048,576-token context, FP8 weights and an Apache 2.0 licence, all listed on its Hugging Face model card.
The same model card sets the hardware floor at a model memory footprint of about 78 GB, with a minimum of 2x A100 80GB, 2x H100 SXM5, or a single H200, B200 or B300.
Beam (Reflection) was announced on 5 October 2026 as a 501B sparse Mixture-of-Experts model with 23B active parameters, trained on 23.8 trillion tokens; the weights are promised under Apache 2.0 later in October 2026 but are not published yet.
The licence separates the two candidates from Meta's Llama weights, which Hugging Face lists under Meta's own community licence rather than Apache 2.0.
Aleph Alpha publishes no public price for its own platform: the weights are free, the two 80GB GPUs that serve them are not.

What is sovereign AI?

Sovereign AI is a claim about control, not a product category: the idea that a company or a country can run, inspect and modify its own models on hardware it governs, instead of renting inference from someone else's API. The releases in October 2026 make that claim checkable in public for the first time, which is why they are worth reading closely.

Three different things get bundled into the phrase, and only one of them is settled by a launch announcement.

Downloadable weights under a permissive licence. Apache 2.0 lets a buyer use, modify and redistribute the model with no user-count cap. This is the part you can verify today, by reading the licence tag on the repository.
Hardware you control. Weights on a disk are not sovereignty if the only machine that can serve them belongs to a vendor you do not control.
Data residency. Where prompts and the KV cache physically sit. That is a hosting decision, and inside the EU it is partly a legal one.

Keep those three separate and both new releases read like a checklist rather than a slogan.

Kolibri: the sovereign AI model you can download today

Aleph Alpha, the Heidelberg lab that has sold on-premise models to European public-sector buyers for years, released Kolibri on 3 October 2026, deliberately on the Day of German Reunification. It is a bilingual German and English Mixture-of-Experts model, and the weights sit on Hugging Face under an Apache 2.0 licence.

The model card is unusually candid, which makes the buying decision easier to make:

78,103,074,560 total parameters with 3.46B active per token, so inference cost tracks the active count while weight storage tracks the total.
1,048,576-token context, though Aleph Alpha recommends staying at or below 262,144 tokens for serving efficiency.
FP8 weights in blocks of 128x128, with embeddings, the LM head, norms and the MoE router kept in bfloat16.
Reasoning mode and tool calling are trained in, with a knowledge cutoff of 18 June 2026.
Hardware floor: roughly 78 GB of model memory, and a minimum of 2x A100 80GB, 2x H100 SXM5, 1x H200, 1x B200 or 1x B300. The card recommends 2x H100 SXM5, 2x H200, 1x B200 or 1x B300.
Training disclosure: 20 trillion pre-training tokens (about 62.5% English, 23.9% German, 13.6% code), then 3.44 trillion mid-training tokens and 201B for long-context extension, on 768 NVIDIA B200 GPUs for 21 days, which the card puts at 392k GPU hours.

Two details matter more than the specs. First, the repository was live on 2 October and had 4,138 downloads and 664 likes by 6 October, with community quantisations (GGUF, MLX 4-bit and 3-bit, an NVFP4 build) appearing inside 24 hours. Quantised builds are what let smaller machines touch a 78B model at all. Second, Aleph Alpha states that it is a signatory of the EU GPAI Code of Practice, and that the design goal was on-premise deployment without sending internal data to a third-party inference service. That is the sovereignty claim in writing, and it comes with an obligation: you supply the GPUs.

Beam: 501B parameters, no weights yet

Reflection announced Beam on 5 October 2026 and calls it the company's first open-weight model: a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, aimed at coding, reasoning and agentic workloads.

The numbers in the announcement are large in a way that is genuinely useful for a buyer weighing self-hosting. Beam was pre-trained on 23.8 trillion tokens, then reinforcement-learned on 10.5K NVIDIA GB300 GPUs for four weeks, generating more than 100 million rollouts with a maximum context length of 256K tokens and roughly 1.3 billion sandboxes for training and grading. Build that in-house and the cost is not a line item, it is a capital project.

What you cannot do is download it. The announcement says Beam is undergoing final red-teaming and evaluations, that early access runs through a waitlist, and that Reflection will release the weights under an Apache 2.0 licence later in October 2026 along with the technical report, model card and the stack for running and deploying it. A search of the Hugging Face API on 6 October 2026 returned no model repository for Reflection. If sovereignty is the requirement this week, Beam is a promise with a date on it, not an option.

The licence is the sovereignty test, not the download

An open-weight model is only as sovereign as its licence allows it to be. This is where the current generation splits, and the split is easy to check: read the licence tag on the Hugging Face repository instead of the launch post.

Apache 2.0: Kolibri 1, Apertus v1.5 from swiss-ai, Qwen3-8B and Mistral Small 3.1 24B all carry apache-2.0 on their repositories as of 6 October 2026. Commercial use, modification and redistribution are permitted without a user threshold.
Vendor community licences: Meta's Llama repositories are tagged llama3.1 and llama3.2, which points at Meta's own community licence rather than Apache 2.0. The weights are downloadable, but the terms are set by the vendor, not by an open-source licence, and a legal team should read them before a sovereign deployment.

Reflection has committed to Apache 2.0 for Beam in writing, which puts it in the permissive column the moment the files land. Until then the column next to it says waitlist.

What it costs to run a sovereign AI model

The weights are the cheap part. A Kolibri deployment starts at two A100 80GB GPUs, or one H200, B200 or B300, because the FP8 weights alone occupy about 78 GB before any KV cache. That floor decides everything else.

Two consequences follow. First, if self-hosting is the plan but the GPUs are not on the balance sheet, the realistic path is renting 80GB-class hardware rather than buying it, and the hourly spread between providers is wide enough to matter over a month of serving: our GPU cloud pricing roundup covers that comparison. Second, the cheap end of hosting does not apply here. A 78B Mixture-of-Experts model does not fit on a budget VPS from RackNerd or Hostinger; those tiers are for the 3B to 8B models, and the honest caveat is that CPU-only serving of even a small model is slow. If that is your target, the best VPS for local LLMs comparison is the relevant page, and local LLM tools covers the software side.

The blunt recommendation: self-host when the data cannot leave your building or your jurisdiction, not to save money. At $0.25 to $8 per million tokens on hosted APIs, a two-A100 box only pays for itself at volumes most teams never reach.

Which sovereign AI model fits your case?

Match the model to the constraint you actually have.

German and English documents, long context, an on-premise mandate: Kolibri 1. It is downloadable today, Apache 2.0, trained to abstain when the answer is not in your context, and the model card is specific enough to size the hardware before you commit.
Coding and agentic work, tolerable to wait: Beam, once the weights land later in October 2026. Put the waitlist request in now and price the hardware after the model card publishes.
EU provenance at a smaller size: the swiss-ai Apertus v1.5 family. We broke down what Apertus opens and what it costs to run when v1.5 shipped.
A permissive small model right now: Qwen3-8B or Mistral Small 3.1 24B, both Apache 2.0, both runnable on a single consumer GPU.
No GPUs and no on-premise requirement: none of this is for you. A hosted API is cheaper and better, and calling that a sovereignty problem when it is not one is the most expensive mistake in this category.

At a glance

Model	Parameters (total / active)	Licence	Weights available now
Kolibri 1 (Aleph Alpha, released 3 Oct 2026)	78B / 3.46B active	Apache 2.0	Yes, on Hugging Face
Beam (Reflection, announced 5 Oct 2026)	501B / 23B active	Apache 2.0 promised	No, early-access waitlist
Apertus v1.5 70B (swiss-ai, released 24 Jul 2026)	70B	Apache 2.0	Yes, on Hugging Face
Qwen3-8B (Qwen)	8B	Apache 2.0	Yes, on Hugging Face
Llama 3.2 1B (Meta)	1B	Meta community licence (llama3.2 tag)	Yes, on Hugging Face

Open-weight model licences and availability, verified 6 October 2026

FAQ

Is sovereign AI free?

The model is, the hardware is not. Kolibri 1 ships under Apache 2.0 with no user cap, but its model card lists a minimum of 2x A100 80GB, 2x H100 SXM5, or a single H200, B200 or B300, with roughly 78 GB of FP8 weights to hold in memory. Aleph Alpha publishes no public pricing for its own platform, so a hosted alternative to self-hosting has to be quoted.

Can you download Beam's weights yet?

No. Reflection's announcement of 5 October 2026 says Beam is undergoing final red-teaming and evaluations, with early access via a waitlist, and that the weights will be released under an Apache 2.0 licence later in October 2026 together with the technical report and model card. A Hugging Face API search on 6 October 2026 returned no Beam repository.

What hardware do you need to run Kolibri?

The Kolibri model card gives a model memory footprint of about 78 GB in FP8 and a minimum configuration of 2x A100 80GB, 2x H100 SXM5, 1x H200, 1x B200 or 1x B300, recommending 2x H100 SXM5, 2x H200, 1x B200 or 1x B300. Community GGUF, MLX 4-bit and NVFP4 quantisations appeared within a day of release and lower the memory floor, at some accuracy cost.