Aleph Alpha announced Kolibri on 2026-10-03, releasing weights for a model with 78B total parameters. One useful test of this release is the cost of a completed task: if reduced computation comes with more retries or human corrections, self-hosting may not be cheaper than the existing service. Downloadable weights make that comparison possible with a model that teams can actually run.Announcement Model card
This article is dated October 4 in Taipei, with a research cutoff of 09:31. It uses the previous day’s release, still within the 48-hour window. NewsMinimalist’s report index displays October 3 at 08:01 UTC, or 16:01 in Taipei; that is the timestamp on the report index. The official announcement does not specify an activation time, so a news timestamp should not be treated as the moment the weights became available.Timing index
Kolibri’s weights come under Apache 2.0, allowing commercial use and self-hosting. Independent dpa reporting describes a focus on administration and industry; the model card specifies German and English. The concrete new option is access to weights, letting teams choose where to run the model and which version to retain. The report does not provide a full cost comparison from customers already using it in production.License and downloads Independent reporting
The mixture-of-experts architecture activates approximately 3.46B parameters per token. Only some experts participate in each computation, but all 78B parameters still need storage. The official model card puts the FP8 weights at roughly 78GB and lists two NVIDIA A100 GPUs with 80GB each as one minimum configuration; two H100 GPUs are among the recommended configurations. Active parameter count explains sparse computation, but cannot by itself establish how many users a server can serve.Architecture and hardware Technical report
Long documents need memory beyond the weights
The model card lists a maximum context of 1,048,576 tokens, while recommending no more than 262,144 tokens for latency- or throughput-sensitive deployments and complex tasks. Those numbers describe capacity and usage guidance respectively, not a promise of equal quality across a million tokens. Processing content also creates a KV cache, which shares GPU memory with the weights. Longer documents and more simultaneous requests both affect available capacity.Context and deployment limits Cache mechanism
The company also presents internal customer-proxy evaluations for areas including automotive suppliers and semiconductors. These are vendor-designed tests, not records of savings achieved after deployment.Application evaluations
I would therefore include Kolibri in a comparison for German-English document products, rather than assume free weights suit every workload. Keeping a model version locally gives a team control over upgrade timing and an opportunity to adapt processing to particular documents. The service still requires maintenance, however, and allocated hardware does not become costless when requests stop. Steady usage and an occasional large batch may favor different ways of obtaining capacity.
Suppose a product processes similarly structured German documents every day. Its team can spread correction time and server costs across a continuing stream of completed work. Occasional complex cases that require repeated checking could produce a different cost outcome on the same hardware and model. This is a hypothetical consequence of the deployment mechanism, not a published customer case. A free license does not settle which workload is economical.
The German-English focus also helps define the scope of adoption. For Taiwanese teams processing Traditional Chinese documents, the sources reviewed here do not supply a local-document quality comparison sufficient to justify replacing an existing model. Applying German-English evaluations or European development provenance to Chinese workflows would overlook differences in language and document type. Open weights enable teams to investigate those differences; they do not supply the result in advance.
The release adds a model that teams can maintain themselves. To turn that option into a product benefit, compare total cost per completed task at the same document type and acceptable quality, including retries, human correction and unused capacity. If those expenses absorb the savings from sparse computation, a lower active parameter count alone does not justify changing services.
The cover reuses this site’s photograph from an NVIDIA European AI infrastructure event. It does not depict Aleph Alpha or the Kolibri launch; the official image download failed because hostname resolution was unavailable.
Sources: