The Vault Doors Open: Beam, Kolibri, and Mistral Large 4 Take Aim at Closed AI

For most of the past two years, a business that wanted a powerful model it could run on its own hardware had one realistic shopping aisle, and nearly every label on it was Chinese. GLM, Qwen, Kimi, DeepSeek, and lately Xiaomi's MiMo set the pace. Western labs mostly kept their best work behind an API. This week that changed, at least on paper.

In the space of four days, Nvidia-backed Reflection AI, Germany's Aleph Alpha, and France's Mistral each announced a new open-weight model. LinkedIn News summed up the pitch all three are making: enterprises and governments want more control, more transparency, and lower costs, and recent incidents of AI systems slipping past safeguards have made that pitch easier to sell.

What actually shipped

Reflection AI: Beam

Reflection, founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, unveiled Beam on October 5. It is a text-only mixture-of-experts model with 501 billion total parameters, of which only 23 billion work on any given word. That small active number is the point: it is what keeps running costs down. Reflection says it trained Beam from scratch on 23.8 trillion tokens, gave it a one-million-token context window, and ran more than 100 million reinforcement-learning attempts on 10,500 Nvidia GB300 chips over four weeks.

The company claims Beam matches Z.ai's GLM-5.2 on advanced reasoning while using three to four times less inference compute than leading Western open models. Two cautions apply. Those numbers are Reflection's own, with no independent testing yet. And you cannot download Beam today: weights, a technical report, and a model card are promised later this month under Apache 2.0, after final red-teaming.

Aleph Alpha: Kolibri

Aleph Alpha released Kolibri on October 3, and unlike the other two, the weights are already on Hugging Face. It is a 78.1-billion-parameter German and English model with just 3.46 billion parameters active per token, small enough to serve on a single H200 or two A100 cards. Apache 2.0 covers the weights. Aleph Alpha built it for European public-sector and regulated buyers who need to know exactly where their model lives.

The modesty is useful. On Aleph Alpha's own tests, Kolibri scores 75.5 overall in English against 80.2 for Alibaba's dense Qwen3.8 27B. Its model card also says a person should review its output before anyone acts on it. That is a sovereignty product, not a frontier contender, and it does not pretend otherwise.

Mistral: Large 4, nicknamed "Le Chonk"

Mistral launched Large 4 on October 6: roughly 1.05 trillion total parameters, 49 billion active, a vision encoder, a one-million-token window, and fluency in more than 160 languages. It is in public preview through Mistral's API now, with open weights promised by the end of October. Mistral calls it the best open-weights model from the United States or Europe.

Independent numbers back that narrow claim. On Artificial Analysis's Intelligence Index, Large 4 Preview scores 38, the highest of any open model built outside China. Nvidia's Nemotron 3 Ultra sits at 23 on the same chart. Context matters, though: Large 4 ties OpenAI's GPT-6 Luna, sits well below Claude Opus 5.5 at 58, and trails China's best open models, led by MiMo-V2.6-Pro at 46, GLM-5.3 at 45, and Kimi K3 at 44.

Where this leaves the scoreboard

• Closed frontier: still clearly ahead. Opus 5.5, Fable 5.1, GPT-6 Astra, and Gemini 4 Argon lead the index by double digits over any open model.
• Chinese open models: still the strongest open tier, roughly 7 to 8 points above the best Western entry.
• Western open models: no longer an afterthought. Mistral leads today, Beam claims parity with GLM-5.2 pending tests, and Axios reports more Western open releases are coming this month.
• Two of the three headline models are announcements, not downloads. Only Kolibri can be run on your own hardware today.

How this could hit closed-model companies

Nobody is going to cancel a Claude or GPT subscription because of a press release. The pressure is slower and more structural, and it shows up in four places.

1. The middle of the price list. Closed labs earn steady volume from their mid-tier models: Sonnet, Sol, Luna. Those are exactly the models an open release at index 38 now matches. When a buyer can get Luna-level results on hardware they control, the mid-tier price has a ceiling it did not have last month. OpenAI's own move this autumn, cutting Sol's API price in half, suggests the labs already feel it.
2. The regulated buyer. Banks, hospitals, defense contractors, and EU agencies have struggled to send sensitive data to an outside API. They also could not easily justify running Chinese weights. A Western, permissively licensed model removes both objections at once, and that is a customer segment closed labs had largely to themselves.
3. The safety argument cuts both ways. After the agent-breach disclosures of the past month, being able to inspect, log, and fence in a model you host yourself is a real selling point. But open weights cannot be recalled. Once Beam or Large 4 is downloaded, safeguards are whatever the person running it decides they are. Closed labs will lean on that argument, and Washington's preference for voluntary safeguards makes it a live political question.
4. The frontier stays closed for now. The hardest agentic work, the longest autonomous runs, and the best writing still come from closed models. Expect closed labs to defend that top tier on capability and push harder on what open weights do not include: managed security, uptime, support, and indemnity.

The governance layer governments could not build on closed models

Price is the visible argument for open weights. The quieter one is governance. A closed model is a rented service: the buyer takes the vendor's word for how it was trained, how it was secured, and what happens to the data inside it. An open model is an artifact the organization holds. It can be inspected, fenced in, and accounted for on the buyer's own terms, which is the only way certain institutions were ever going to touch AI at all.

This pattern is established practice, not a prediction. Meta opened Llama to US government agencies and defense contractors in 2024 because, as the company put it, an open model can be securely downloaded and deployed without sensitive data passing through a third-party provider, fine-tuned on classified data, and hosted in secure environments at various classification levels. Scale AI built Defense Llama from Llama 3 for that audience and has shown it running on classified networks for combat planning and intelligence work. Meta has since extended access to Five Eyes partners, and in September 2025 to allies including France, Germany, Italy, Japan, and South Korea, plus NATO and EU institutions. The Defense Department's AI Rapid Capabilities Cell, formed by the Chief Digital and AI Office and the Defense Innovation Unit, exists to scale exactly these tools across the department.

European public institutions build the same way. Germany's federal government runs PLAIN, a sovereign data and AI platform operated by Bundesdruckerei, so federal ministries can do AI work inside a protected national environment. Aleph Alpha built Kolibri for precisely this buyer: small enough to serve on two data-center cards, licensed Apache 2.0, weights already published. Research on sovereign deployments for regulated industries spells out what the governance layer looks like in practice: zero data egress, air-gapped operation, retrieval grounded with enforced citations, and cryptographically signed update chains so an auditor can verify what the system is actually running.

Regulators have started treating the two model classes differently. Under the EU AI Act, providers of general-purpose models released fully open source, with parameters public and no monetization, are exempt from several obligations, including maintaining technical documentation for authorities and appointing an EU representative. The exemption has hard limits: systemic-risk models must comply regardless, and the copyright policy and training-data summary requirements still apply. Closed-model providers owe the full documentation package to the EU AI Office. Europe's rulebook already treats openness as a form of verifiability.

The honest version of the argument includes the counterweight. Open weights shift the governance burden onto the buyer: free weights still need monitoring, patching, access control, and someone accountable when the output is wrong, while a closed vendor bundles those controls into the contract. US policy review is active; the Commerce Department's NTIA has examined whether to restrict wide availability of model weights, including licensing regimes and staged release. And a downloaded model cannot be recalled, so whoever runs it decides what safeguards exist. Open weights enable a governance layer closed models cannot offer where data legally cannot leave the building. For everything else, the buyer is choosing which problems they would rather own.

The layer underneath got its own open release the same day

Frontier chat models are only half of a working AI system. The other half is the embedding model, the piece that converts documents, images, and audio into vectors so search and retrieval-augmented generation can find the right material before a model answers on it. Google joined the open wave the same day as this week's announcements. EmbeddingGemma 2 is a 740-million-parameter model that maps text, code, images, video frames, and audio into one shared 768-dimensional space, built on the Gemma 4 architecture and released under Apache 2.0. It is the first small, natively multimodal embedder built for consumer hardware rather than the data center.

The interesting part is the modularity. The text model is 270 million parameters, with a 170-million-parameter vision encoder and a 300-million-parameter audio encoder you load only if your task needs them. Google reports the text-only weights need about 191MB of RAM on a Pixel 11 Pro, and the full multimodal model about 567MB. Matryoshka training lets you truncate vectors from 768 down to 128 dimensions, cutting vector storage by up to 6x with minimal quality loss. It handles more than 100 languages, roughly a 14% improvement on code tasks over its predecessor, and an 8K-token context window that fits minutes of audio or video. That predecessor, a text-only 308-million-parameter model released in September 2025, has been downloaded more than 20 million times, which tells you how much demand exists for embeddings that never leave the device.

This matters for the argument above because retrieval is where the sensitive data actually sits. An organization can now assemble a complete local stack with no component owned by an outside provider: an open embedding model such as EmbeddingGemma 2 to index its own material, and an open language model such as Kolibri to answer from it. Zero data egress at every layer, on a single server or even a laptop. Google's hosted twin, Gemini Embedding 2, offers the same multimodal approach as a managed API for teams that do not need to self-host. The open wave now covers the generation layer, the retrieval layer, and the policy layer around both. What used to require a cloud contract at every step can run inside your walls.

What a small team should do with this

• Do not rebuild anything yet. Beam and Large 4 weights are not out, and independent evaluations are still coming. Revisit at the end of October.
• List your AI tasks by sensitivity. Anything touching client data, contracts, or regulated information is the first candidate for a self-hosted or private-cloud open model.
• Read the license, not the headline. Apache 2.0 and MIT are simple. Custom licenses often carry conditions for large commercial use.
• Plan the retrieval layer alongside the model. An open embedding model such as EmbeddingGemma 2 paired with an open language model gets you a private search stack where no document ever leaves your machine.
• Budget for the parts that are not free. Free weights still need GPUs, monitoring, patching, and someone accountable when output is wrong.
• Write down who approves a model change. Swapping models is now cheap; swapping them without review is how governance quietly erodes.

None of the three models beat the closed frontier this week. What changed is slower and more durable: the closed labs' comfortable middle ground now has Western-made competition at every layer, from chat models to the retrieval and governance stack around them, and buyers who wanted neither a premium API nor Chinese weights finally have a third option. Kolibri and EmbeddingGemma 2 are the two you can download today, which makes them the ones worth testing first. Watch what happens to mid-tier pricing when the Beam and Large 4 weights land by the end of October.