Three Open-Weight Launches in Five Days — and They're Not Fighting the Same Battle
October 6, 2026 · 5 min read
Between October 1 and October 5, three serious open-weight releases landed within five days of each other. Nvidia-backed Reflection shipped Beam, a 501-billion-parameter MoE. Germany's Aleph Alpha released Kolibri, a 78.1-billion-parameter MoE built for European sovereignty. And the Allen Institute for AI dropped Olmo-core 3, a training framework that makes big MoE models dramatically cheaper to train. Same week, same "open" label — but they're solving three completely different problems.
That is the interesting part. Two years ago, every open model was playing the same game: more parameters, bigger context window, top the leaderboard. Now the front has split. One contender is trying to out-efficiency the Chinese labs, another is selling sovereignty to European governments, and a third isn't even selling a model — it's selling the factory that makes models cheaper. Open-weight competition just stopped being one race.
Beam: punching up at GLM-5.2, at a quarter of the compute
Reflection, an Nvidia-backed startup, unveiled its first AI model on October 5, explicitly positioning it against Chinese open models. Beam packs 501 billion parameters as a mixture-of-experts, but only activates 23 billion of them per inference pass — the MoE trick of looking huge while running small. It supports a million-token context window. The headline claim: reasoning benchmark performance on par with Zhipu's GLM-5.2, at one-quarter to one-third of the inference compute. Weights drop under Apache 2.0 later this month. (via Reuters)
Kolibri: the European sovereignty play
Two days earlier, on October 3, Germany's Aleph Alpha released Kolibri — a 78.1-billion-parameter MoE that activates just 3.46 billion parameters per token, with a context window stretching up to a million tokens. The weights are already on Hugging Face under Apache 2.0. Kolibri's pitch is not raw capability, though: it's sovereignty. Trained in Germany and Finland, compliant with the EU AI Act and GDPR, bilingual in German and English, and deployable on-premises. This is a model designed to be bought by European governments and enterprises that cannot, legally or politically, ship their data to someone else's cloud. (via runtimewire)
Olmo-core 3: attacking the training bill instead
On October 1, the Allen Institute for AI released Olmo-core 3, an open training framework that switches MoE training from FSDP to DDP architecture — and claims it can scale MoE training to trillion-parameter scale with almost no increase in compute cost. The reference build, a 47-billion-parameter model with 128 experts activating 3.2 billion per token, loses less than 5% throughput. On eight B300 cards, each card pushes 52,000 tokens per second — 2.7 times the previous implementation. Olmo-core 3 isn't a product you download and chat with; it's the machinery behind the products. It says the next leap in open models won't come from cleverer architectures alone — it will come from making giant training runs affordable to more labs. (via completeaitraining)
One "open" label, three different races
Put them side by side and the split is stark. Beam's bet: don't out-muscle the Chinese labs — out-efficiency them. If a model matches GLM-5.2 at a quarter of the inference cost, the raw parameter race stops mattering. Kolibri's bet: capability was never the whole sale — compliance, local deployment, and a European passport are. It doesn't need to beat Beam on benchmarks; it needs to be the model a German hospital or ministry can actually use. And Olmo-core 3's bet sits underneath both: if training a trillion-parameter MoE costs roughly what a 47-billion one used to, the entire cost curve of open-weight AI shifts. That's the maturation sign. Open-weight competition no longer means everyone piling parameters onto the same pile. It means picking your battlefield — efficiency, sovereignty, or infrastructure — and letting the others fight theirs. The next twelve months of open models will be shaped less by who builds the biggest model than by which of these three bets pays off first.