Kolibri-1 hardware requirements: GPUs, VRAM and users for Aleph Alpha's open model

Kolibri-1 on one GPU or two

Kolibri-1, Aleph Alpha's open German-English model, needs about 73 GiB of GPU memory for its FP8 weights, so it fits one 96 GB card or one 141 GB card. By our users-per-card rule, one NVIDIA RTX PRO 6000 Blackwell Server Edition holds the model with about 28 concurrent conversations of 32K tokens, two Server Edition cards split the model with about 279, and one H200 NVL about 150. For a service used by a department, two cards or one H200 NVL are the configurations we would size first.

Aleph Alpha GmbH published Kolibri-1, an open-weight model for German and English, on Hugging Face on 3 October 2026. Its model card states 78,103,074,560 parameters in total and 3,457,573,120 active per token. It suits organisations that want a model with strong German on their own hardware, and our guide to open LLMs for European languages places it among EuroLLM, Teuken and the others.

Aleph Alpha, Kolibri-1 model card, config.json and file list on Hugging Face, read on 24 October 2026. 78.8 GB is the sum of the 32 safetensors files, 78,841,467,208 bytes, or 73.4 GiB.

The card recommends at most 262,144 tokens "for serving efficiency and complex tasks", and the longer window up to 1,048,576 tokens needs extra vLLM flags. Embeddings, the output head, norms and the expert router stay in BF16. A BF16 repository, Kolibri-1-BF16, is the base model of the FP8 release and is 156.2 GB.

What Aleph Alpha lists as minimum hardware

The model card reads: "Minimum: 2× A100 80 GB, 2× H100 SXM5, 1× H200, 1× B200 or 1× B300." Recommended are two H100 SXM5, two H200, one B200 or one B300. Every entry on that list has at least 141 GB in one card or 160 GB in two, about twice the weights, so Aleph Alpha plans for room beyond the model itself.

The model card does not mention RTX cards at all. Of the cards on the list, the H200 is the one we supply: the NVIDIA H200 NVL is its PCIe version, with the same 141 GB of HBM3e and the same Hopper architecture, so it is the closest match to the "1× H200" line. The RTX PRO 6000 Server Edition has 96 GB per card, between the 80 GB of an A100 and the 141 GB of an H200, and a different chip generation, compute capability 12.0. Whether it runs Kolibri-1 depends on the software, which we cover below.

How much KV cache one conversation needs

Kolibri-1 needs little memory per conversation. Its config.json repeats a pattern of four sliding-window layers and one full-attention layer, so only 10 of the 50 layers keep every token of the context. The other 40 keep a window of about 512 tokens (513 in config.json). Each layer stores keys and values for four KV heads of dimension 128, which is 1 KiB per token in FP8.

The serve command on the model card sets --kv-cache-dtype fp8, and Aleph Alpha states it evaluated the model with an FP8 KV cache. vLLM's design notes on its hybrid KV cache manager (page of 13 October 2026) describe how sliding-window layers free the blocks that fall outside the window, while full-attention layers keep a slot for every token.

Our estimate from config.json: 10 full-attention layers × 2 × 4 × 128 values per token, plus 40 sliding-window layers × 513 tokens. If an engine kept the full context in all 50 layers, the FP8 figure would be 1.56 GiB at 32K.

For comparison, Llama 3.3 70B needs 5 GiB of FP8 cache for the same 32K conversation, as our guide to how many users one RTX PRO 6000 can serve works out. With Kolibri-1 the weights decide the card count; the conversations add little.

Conversations per configuration

Our users-per-card rule takes 90 per cent of the memory the driver reports and subtracts 3 GiB per card: 83.0 GiB per RTX PRO 6000 Server Edition, 123.4 GiB per H200 NVL and 25.7 GiB per RTX PRO 4500 Server Edition. The weights come off that budget once per copy, and the rest holds KV cache.

Our estimate by the users-per-card rule: concurrent conversations at full length with an FP8 KV cache, weights 73.4 GiB per copy. Requests in flight at the same moment, not staff using the service.

With an engine that kept the full context in every layer, one Server Edition card would hold 6 conversations of 32K, one H200 NVL 31 and a TP=2 pair 59: still enough for a team, which is why we size with both figures until the delivered system shows which one applies.

Two copies on two cards hold fewer conversations than one copy split over both, and they keep working when one card fails or one engine restarts. With four KV heads, tensor parallelism splits the cache evenly over two or four cards. The two cards exchange data over PCIe, and for a pair vLLM keeps its own all-reduce kernel, as our article on RTX PRO 6000 without NVLink explains.

A single card at the full 262,144 tokens holds three conversations by this rule. Long-document work at that length belongs on two cards or one H200 NVL. We supply servers with one, two or four RTX PRO 6000 Server Edition cards and with H200 NVL; send us your expected context length and peak requests, and our team replies with a configuration for Kolibri-1.

Speed: 3.46B active parameters per token

Generating one token reads the active weights once. For Kolibri-1 that is about 2.8 GB of FP8 expert and attention weights plus the 0.66 GB output head in BF16, about 3.5 GB in all by our estimate. Divided into the memory bandwidth, that gives an upper bound by our estimate of about 450 tokens per second per stream on one RTX PRO 6000 Server Edition at 1,597 GB/s, and about 1,370 on one H200 NVL at 4,800 GB/s.

Real figures are lower: the KV cache is read as well, kernels do not reach the full bandwidth, and batched requests route to different experts, so a batch reads more of the 384 experts than one stream does. We found no published measurement of Kolibri-1 on either card as of 24 October 2026. The ceiling still shows the point for buyers: with 3.46B active parameters, Kolibri-1 answers fast on any of these cards, and the choice between them comes down to memory, software support and the number of conversations.

Kolibri-1 in vLLM on the RTX PRO 6000: plugin and FP8 blocks

Three things decide whether Kolibri-1 runs well on a Blackwell RTX card.

1. The vLLM plugin. The model card says Kolibri-1 "requires the aleph-alpha-inference package that provides the Kolibri vLLM plugin", installed with pip install 'aleph-alpha-inference>=1' or as the container image ghcr.io/aleph-alpha/aleph-alpha-inference. The package brings its own supported vLLM version, which the card does not name. The model type in config.json is kolibri1.
2. FP8 in 128 × 128 blocks. The weights carry one scale per block of 128 × 128 values, with activations quantised at run time. Our guide to Blackwell RTX (SM120) software support shows the state as of 24 October 2026: vLLM 0.30.0 lists an SM12x blockwise FP8 kernel, SGLang names a CUTLASS block-FP8 backend for SM120, and TensorRT-LLM lists no FP8 block scaling for sm120. Which kernel the plugin's vLLM build selects on this card is not documented.
3. Hopper as the safe path. The H200 NVL is compute capability 9.0, and TensorRT-LLM's matrix lists FP8 block scaling for Hopper; one H200 is on Aleph Alpha's list.

Transformers supports the model natively from version 5.20, a fallback for tests, not for serving many users. Before a Server Edition system goes into service, run the plugin's container on it with your prompts and read the start-up log for fallback warnings. Our guide to GPU server acceptance testing lists what to check on delivery.

Which configuration to choose

For a director, there are three sensible options.

• One RTX PRO 6000 Server Edition for a pilot: the model fits, and about 28 conversations of 32K by our rule are enough to test German quality on your own documents. Plan the software check above.
• Two RTX PRO 6000 Server Edition cards for a department or a public body's first service: two copies for resilience, or one TP=2 copy for long contexts. Our page on the two-card Server Edition server has the configuration; we estimate about 1.8 kW at the wall with two AMD EPYC 9135 processors.
• One or two H200 NVL when you want the hardware on Aleph Alpha's list and the higher bandwidth. A two-card server with an NVLink bridge draws about 2.0 kW with two AMD EPYC 9355 by the estimate on our two H200 NVL server page.

Four RTX PRO 4500 Server Edition cards suit a server that also hosts virtual desktops: the card is on NVIDIA's vGPU list from release 20.0, and four of them hold Kolibri-1 as one TP=4 copy with 88 conversations of 32K by our rule. Our article on the four-card RTX PRO 4500 Server Edition server in 2U covers that chassis. Sizing a server for a whole administration is in our guide to GPU servers for public administration.

Licence, languages and what the model card reports

The LICENSE file is the Apache License 2.0 with no added clauses (11,358 bytes, the size of the standard text), and the model card states that Aleph Alpha does "not implement an acceptable use policy". Its section on responsible use asks users not to use the model for practices prohibited under Article 5 of the AI Act or for copyright infringement, asks for use "following ethical standards" and advises to "avoid unsupervised use in high-stakes environments". These are requests on the model card, not added licence terms. Aleph Alpha states that the licence covers the published weights and configuration files and does not extend to its training code or methods. The knowledge cutoff is 18 June 2026 for both languages.

Aleph Alpha reports an overall benchmark score of 75.5 in English and 70.8 in German for the post-trained model at high reasoning effort, and 96.9 on AIME 2025 in English. These are the vendor's own figures; test the model on your own German documents before you size a service on it.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 6000 Server Edition and the H200 NVL as cards and in AI servers built to order, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component; we install the operating system, drivers, CUDA and a container runtime on request. Setting up Kolibri-1, retrieval on your documents and the platform around it is our Private AI/ML service.

FAQ