Skip to content
VECTOR WIREAI INTELLIGENCE
UTC
Home

Beam and Kolibri bet Western open weights can undercut closed AI on cost

Reflection AI released Beam, a 501-billion-parameter open-weight model that the company says matches leading Chinese systems on reasoning benchmarks at a fraction of the compute cost1. The launch arrives days after Aleph Alpha shipped its own sovereignty-focused model and amid reports that still more Western open-weight releases are imminent4,7.

Why it matters

The open-weight tier has been dominated by Chinese labs, leaving Western enterprises that wanted to self-host powerful models with few domestic options. That is changing fast. Reflection AI released Beam on Monday, describing it as a text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, pre-trained on 23.8 trillion tokens with a one-million-token context window. The company says Beam scores on par with Z.ai's GLM-5.2 on advanced reasoning benchmarks and "outperforms today's leading Western open models while using '3-4x less inference compute'". Aleph Alpha has already shipped Kolibri, a 78.1-billion-parameter Mixture-of-Experts model under the Apache 2.0 license, purpose-built for European sovereignty requirements. And Axios reports that Reflection's launch "is set to be joined by other open-weight models from other Western players this month". Taken together, the wave marks the first time multiple well-resourced Western entrants have converged on the open-weight segment simultaneously, challenging both the closed-model pricing moat and the assumption that open-weight leadership belongs to Chinese labs.

The big picture

Reflection AI, founded by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, is positioning Beam as a "workhorse model" for enterprises, the public sector, and developers5. CEO Laskin told Semafor the model "is particularly strong at coding and agentic tasks" and "needs three to four times less computing power to reason through a problem than comparable open models, making it faster and cheaper to run". Beam's capabilities are "comparable to" Z.ai's GLM-5.2, released in June, and "approaching" the performance of Alibaba's Qwen 3.8-Max, which itself trails the top closed American models from OpenAI and Anthropic. For context, Z.ai's GLM-5.2 has roughly 744 billion total parameters with 40 billion active, compared with Beam's 501 billion total and 23 billion active.

Reflection's last funding round valued the company at a $25 billion pre-money valuation, and the Nvidia-backed startup has signed deals with Nebius and SpaceX to rent Nvidia AI servers. It has also begun testing a sovereign AI factory partnership with Shinsegae Group in South Korea. The strategic logic is explicit: "Combining a lower-tier open model with high-quality company data can yield results that rival the most expensive frontier AI systems in certain situations".

Aleph Alpha's Kolibri occupies a complementary niche. The model carries 78.1 billion total parameters but activates only 3.46 billion per token, supports contexts up to one million tokens, and was trained on 768 B200 GPUs over nearly 24 trillion tokens. It is bilingual in English and German, with German accounting for 21.3% of pre-training tokens and a 128,000-entry vocabulary designed to preserve German compound words.

ANALYSIS The two launches reveal distinct but converging strategies. Reflection is coupling an open model with Nvidia hardware partnerships, packaging compute and weights together as an "AI factory" product that institutions can localize. Aleph Alpha is selling regulatory geography: a model whose entire provenance chain sits inside Europe, aimed at "government and regulated industries". Both strategies treat the model itself as a commodity layer and compete on what surrounds it: data residency, hardware access, and compliance guarantees.

Reflection's performance claims have not been independently verified. Semafor's framing is notably hedged: Beam is "approaching" Qwen 3.8-Max, not matching it, and Qwen 3.8-Max itself "trails the top closed American models". That double qualifier places Beam at least two rungs below the frontier, a gap the company's cost argument must bridge. The bet is that most enterprise workloads do not require frontier capability, only sufficient capability at manageable cost.

Aleph Alpha's architecture choices reinforce the cost argument. Kolibri uses 384 experts with six active per token, full attention in 10 of 50 layers, and a 512-token sliding window in the remaining 40 to contain inference costs. On AIME 2025, Kolibri scored 96.9; on LiveCodeBench v6, 85.9.

For closed-model providers, the threat is less about raw capability than about procurement math. If enterprises can fine-tune an open-weight model on proprietary data and run it on-premises for a predictable hardware cost, the per-token API fees charged by Anthropic, OpenAI, and Google face downward pressure in exactly the enterprise segments where margins are highest.

Open-weight models also carry regulatory tension. Axios notes that such models "can be harder to monitor and regulate than the Big Three so-called frontier models," while "supporters of such ecosystems say their widespread availability and transparency also bring security advantages". That unresolved debate will intensify as more capable open-weight systems enter circulation.

What's next

Reflection says it is already training its next model. Additional Western open-weight launches are anticipated this month. Aleph Alpha's Kolibri is already downloadable from Hugging Face9. Aleph Alpha says it has built sector-specific evaluation suites for public administration, automotive, semiconductors, industrial technology, and aerospace, a signal that the company is already selling into procurement cycles where benchmarks alone do not close deals.

Was this useful?

The day’s vector, in your inbox.

Vector Wire Daily: one dispatch each publishing day.