Beam 1.0 vs Kolibri: Two Apache 2.0 Releases Are Not the Same Release

Beam 1.0 and Kolibri are the two most comparable open-weight releases of the last month, and comparing them is more useful than comparing either to a frontier API model, because both can actually be downloaded. Aleph Alpha's Kolibri arrived on 3 October 2026 — 78,103,074,560 total parameters, 3,457,573,120 active per token, a validated context window of 1,048,576 tokens, German and English, Apache 2.0. Reflection's Beam 1.0 arrived on 16 October 2026 with 501.3 billion total parameters, 23.4 billion active, 1M tokens of context, English, and the same licence. Same licence, similar sparsity philosophy, and almost nothing else in common — which is exactly why the pairing tells you more than a size-versus-size comparison would. One was built by a lab with a jurisdiction argument to make; the other by a lab with a capability argument to make.

The two also differ in a way that decides most real deployments before a single benchmark row is read: what the vendor published alongside the weights. Aleph Alpha shipped an unusually complete account of how the model was trained — data provenance, energy, the failed predecessor, the exact GPU-hours. Reflection shipped a verifier you can point at your own endpoint. Those are different kinds of trust, and which one your procurement process accepts is the first question to answer.

Same licence, different products

Start with what is genuinely identical, because it is less than it looks.

• Licence — both Apache 2.0, both with the licence in the repository rather than in a blog sentence. Beam's card adds that distillation, fine-tuning, quantisation and merging are not restricted. Kolibri carries no acceptable-use rider, no monthly-active-user threshold and no separate commercial terms.

• Sparsity strategy — both are mixtures of experts with aggressive activation ratios. Kolibri wakes 3.46B of 78.1B, roughly 4.4 per cent; Beam wakes 23.4B of 501.3B, roughly 4.7 per cent. Neither lab is selling you a dense model.

• Long context — Kolibri states a 262,144-token native window with quality and serving efficiency validated to 1,048,576; Beam states 1M on the card and 131,072 tokens of maximum output on the API.

• No independent benchmark reproduction — neither vendor's table has been checked by a third party. Artificial Analysis has no model page for Kolibri at all, while it does list Beam 1.0 at an Intelligence Index of 30, nineteenth of 120 on index version v4.3.2.

That last line is the first real asymmetry and it is easy to miss. Beam has one outside number; Kolibri has none. A 30 on an index that covers 120 models is not a flattering figure, but it is a figure that Reflection did not produce, and for a buyer deciding whether a vendor's table is trustworthy, the existence of any independent measurement is worth more than the value of it.

What each vendor chose to publish

Aleph Alpha's release documentation is the more unusual artefact. Alongside the weights it published 20 trillion pre-training tokens drawn from a filtered bilingual corpus — roughly 62.5 per cent English, 23.9 per cent German, 13.6 per cent code — plus 3.44 trillion mid-training tokens and 201 billion for long-context extension. Compute is stated as 768 N​VIDIA B200s across 96 HGX nodes, 21 days of pre-training for 511 hours and 392,000 GPU-hours at a reported 6.4×10²³ FLOPs, with mid-training adding five days and 90,000 GPU-hours and long-context extension adding 13 hours and 10,000 GPU-hours. Energy is estimated at 9.5×10² MWh including data-centre overhead, covering pre-training, mid-training and long-context training, and excludes supervised fine-tuning and reinforcement learning.

Two things follow from publishing that. First, an organisation with a procurement rule that starts with "where did the training data come from and under which law" has an answer rather than a conversation. Second, and less comfortably, the same documentation shows the model losing to Qwen3.8 27B on Aleph Alpha's own harness — 75.5 against 80.2 on the English average and 70.8 against 79.9 on the German average, with Q​wen ahead on GPQA Diamond, LiveCodeBench v6, SWE-Bench Verified and both long-context benchmarks. Gemma 4 26B-A4B IT sits at 71.9 and 66.3 on the two averages. Kolibri's case is not that it is the strongest model of its size; it is that it is the strongest model of its size whose training you can audit. That documentation is the reason the release reads as research rather than as marketing: Aleph Alpha is describing how the model was made because it wants the pipeline, not the model, to be the product, and a lab selling a training pipeline publishes the numbers that show the pipeline worked.

Reflection published a different kind of thing. Beam 1.0's release includes reference serving configurations, a Helm chart, day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and N​VIDIA Dynamo, and the beam-verifier CLI, which runs a conformance suite against a live endpoint and then evaluates it on the main benchmarks so a deployment can be checked against Reflection's reference scores. What it does not include is a rate card: the documentation carries a full page on rate limits and no per-million-token price, and the announcement says only that the API is free to start.

What it costs to run each one

Neither is served on our catalogue, and we are not going to imply otherwise. Both are self-hosted purchases, and the hardware floors are the real price.

• Beam 1.0 — the NVFP4 build is 273 GiB and the reference configuration is four N​VIDIA B300 SXM GPUs on NVLink, a single box. FP8 is 468.72 GiB and runs on eight H200s; BF16 is 933.81 GiB and lands on eight B200s or B300s. A 1.93B DFlash2 drafter ships alongside for speculative decoding, worth 1.7× to 3.0× decode speedup on the NVFP4 checkpoint. The serving recipes keep the KV cache in FP8 E4M3 with a 64-token page.

• Kolibri — the FP8 checkpoint is the documented serving precision, with 128×128 block weights and dynamically quantized activations, evaluated with an FP8 KV cache while embeddings, the LM head, the norms and the MoE router stay in bfloat16. A vLLM tool-call parser ships in the same repository as the weights, and the repository is ungated — an Apache 2.0 download rather than a request.

The comparison a buyer should actually run is not benchmark rows. It is a two-GPU self-hosted endpoint that answers in German and English and whose training you can document, against a four-GPU endpoint that is stronger on coding and agentic work and whose training documentation, as far as we can find, does not extend to GPU-hours or energy.

If your workload is regulated documents that cannot leave the building, the first is the only one that closes the argument. If your workload is an engineering agent, the second is the only one with a verifier.

If you want the capability without the rack

Both models describe a class — small-active-parameter MoE with a permissive licence — that is already served on routes you can call today, and pricing the alternative is the fastest way to decide whether owning the hardware is worth it.

• Gemma 4 26B-A4B IT — $0.06 per million input tokens and $0.33 per million output with a 262,144-token window. The nearest thing on a routed endpoint to the Kolibri tier, and the model that actually beats Kolibri on Aleph Alpha's own English and German averages.

• Qwen3.8-27B — $0.33 input and $2.40 output, 262,144-token context, and Apache-licensed open weights if you would rather run them than rent them.

• GLM-5.2 — $1.40 and $4.40 with a 1,000,000-token window, for the long-context and coding work Beam 1.0 is aimed at.

• Kimi K3 — $3.00 input, $0.30 cached, $15.00 output at 1,048,576 tokens, the model Reflection's own launch post concedes is ahead on raw capability.

OrcaRouter routes that set behind one OpenAI-compatible key with the provider's list price passed through and nothing added, so a vendor price cut lands on your bill the same day rather than at the next contract renewal, and automatic failover covers the case where one provider degrades under load. A useful pattern for anyone weighing a sovereign deployment is to run the regulated traffic against a self-hosted Kolibri endpoint and the general work against a routed tier, with a routing rule deciding which is which — one key, one config, and a hardware decision that only has to be justified for the documents that need it.

Which release you were actually looking for

If you searched for open weights because you need to answer a question about provenance, jurisdiction or auditability, Kolibri is the release with the answer. It is not the strongest model at its size — Aleph Alpha's own tables say so — and it is two languages wide, and its recommended serving configuration is two H100 SXM5s or two H200s — or a single B200 or B300, which the hardware table lists alongside them. The trade is explicit and, for a regulated deployment, probably correct.

If you searched for open weights because you want an engineering agent you can run yourself, Beam 1.0 is the release with the capability, the serving stack and the verification tooling, and it asks for four B300s and a price list that has not been written yet. The trade there is capital and ambiguity for a model that does frontier-adjacent coding work on your own hardware.

What the pairing shows is that "open weight" stopped being a differentiator some time ago. Two Apache 2.0 releases from two labs in two weeks can share a licence, a sparsity philosophy and a context window, and still be answers to completely different questions. Sorting by licence gets you nowhere; sorting by which question the release was built to answer gets you to the right folder.

None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.