Aleph Alpha Kolibri: Open-Weight Specs, Benchmarks and Setup

Aleph Alpha released Kolibri on 3 October 2026, the Day of German Unity. It is a bilingual English-German mixture-of-experts model: 78.1B parameters, 3.46B active per token, a validated context of 1,048,576 tokens, and Apache 2.0 weights on Hugging Face as Aleph-Alpha/Kolibri-1.

Kolibri is not on RouterPlex yet. The announcement and the model card publish weights and a vLLM command. They do not publish a hosted API price or a RouterPlex model ID. Create a RouterPlex account if you want a prepaid key ready the day a catalog row exists. Until then, the way to run Kolibri is the open weights.

Kolibri at a glance #

There is no public per-token Kolibri API price. Aleph Alpha's contact for enterprise deployment is their sales page. On-prem GPU time is the bill until a hosted row exists.

What the open-weight model actually is #

Kolibri is the second model through Aleph Alpha's Model Factory. The first, Kolibri Origin, finished pre-training on 11 June 2026 at 30.6B total and 3.27B active, with a 65,536-token trained context. Origin was not released. Kolibri finished pre-training on 11 September and shipped three weeks later.

The size choice is a serving decision, not a "bigger is better" one. Aleph Alpha says a 123B variant could serve only 3 concurrent 256k queries on two H100s, while 78B handled 18 and decoded 28% faster. They also preferred 384 narrower experts over fewer wide ones. Of 50 layers, 10 use full attention (every 5th layer) and 40 use a 512-token sliding window, so decode cost in those layers does not grow with context.

Pre-training ran on 768 B200 GPUs for 21 days and 20T tokens at a 16k sequence length, then 3.44T tokens of mid-training at 64k, then 201B tokens of long-context adaptation at 256k (the blog rounds that last stage to 200B). The model card puts pre-training at 392,000 GPU-hours, mid-training at 90,000, and long context at 10,000. Estimated training energy, including data-center overhead, is 950 MWh. That figure excludes supervised fine-tuning and reinforcement learning.

Routing during training uses exact quantile balancing. Aleph Alpha contrasts that with Kimi K3, which estimates the global quantile from histograms. They say the exact version has a fixed communication cost and improved both load balance and quality. The optimizer is Muon, as it was for Origin. They report no loss spikes on either run.

Post-training is supervised fine-tuning, then reinforcement learning. The SFT mix is 268B tokens after filtering, including 174B of synthetic tokens before that filter. RL used more than 1.2 million tasks. Reasoning effort is a trained control: none, low, medium, or high. The chat template treats an omitted effort as high.

Kolibri vs Kolibri Origin #

Active parameters barely moved. Total parameters, expert count, context, and the German data share are what changed.

Benchmarks Aleph Alpha published #

These numbers come from Aleph Alpha's own harness, with each model at its highest reasoning effort. They are not an independent re-run.

The announcement's comparison set, on a shared 0–100 scale. Higher is better. A dash means Aleph Alpha did not publish a score.

Qwen3.6-35B-A3B is the fair peer: about 3B active, same order of magnitude as Kolibri's 3.46B. Nemotron 3 Super is the "up to four times the active parameters" model Aleph Alpha names (about 12B active). Mistral Small 4 is about 6B active.

Where Kolibri leads that peer set

• Math, including German. AIME 2025 is 96.9, against 91.7 for Nemotron 3 Super and 84.6 for Qwen3.6. German AIME 2025 is 87.5, against 85.6 and 82.9.
• Science questions. GPQA Diamond is 84.3 in English and 81.3 in German. The German drop from the English score is 3.0 points. Origin dropped 9.6.
• One agent bench by a wide margin. τ³-bench banking is 38.1, against 15.5 for Nemotron 3 Super and 10.6 for Qwen3.6. That is still a low absolute score.
• LiveCodeBench v6. 85.9, against 82.5 and 82.0.

Aleph Alpha's overall row, which averages a wider set than the table above, puts Kolibri at 75.5 English and 70.8 German. Among the mixture-of-experts models in that row, those are the high scores. Qwen3.5-35B-A3B is close on English at 74.7. The dense Qwen3.8 27B is higher still, at 80.2 and 79.9, and it activates about 27B parameters per token rather than 3.46B. Aleph Alpha greys dense models out of the "best MoE" marks for that reason.

Where Kolibri trails

• SWE-bench Verified: 66.4. Qwen3.6 is 73.8. Qwen3.5 is 71.6. Qwen3.8 27B is 72.6. Nemotron 3 Super is 60.2.
• Tool calling. BFCL v4 overall is 61.4, behind Qwen3.6 at 67.2. Kolibri is level with Nemotron 3 Super at 61.0.
• Long context. LongBench Pro is 64.5, behind Qwen3.6 at 70.8 and ahead of Nemotron 3 Super at 62.9.
• Terminal work. TerminalBench 2.1 is 27.7. Qwen3.5 and Nemotron 3 Super are both 39.7. Qwen3.8 27B is 76.8. Qwen3.6 has no published cell in this row.
• The omniscience index. −32.8 is better than Origin's −64.0 and worse than Qwen3.6's −15.3. Higher is better, and a negative index means the model is still punished for wrong answers.

If the job is a software-engineering agent or a tool-calling loop, this table does not make Kolibri the default. If the job is German and English math, science questions, or the banking tool bench they published, it does.

German, on purpose #

The announcement says 21.3% of the 20T pre-training tokens are German, about 4.3T, against roughly 62% English and 14% code. The model card describes that same 20T mix as about 62.5% English, 23.9% German, and 13.6% code. Both put German in the low twenties, not in the leftover bin. Each unique German token was seen about 1.8 times.

Open German datasets were not enough. After dedup and filtering, Aleph Alpha had 390B German tokens against a target near 4T. They filled the gap with a German Common Crawl pipeline (1.3T unique tokens; English word-length filters were dropping administrative German), LLM rephrases of German documents (about 1T; the source stays German), and a small translated share. Translation was a Kolibri Origin tool. They argue it imports English-web geography and institutions along with the words.

The tokenizer is their own 128k English-German UniBPE. It is the part of the release that changes your bill if you serve the weights yourself: fewer German tokens means less decode for the same document.

On German web text (FineWeb-2), Kolibri scores 4.90 bytes per token, the best figure in Aleph Alpha's comparison. Their own plain BPE 128k control, trained on the same data, scores 4.89. The compression gain from the training method is a hundredth of a byte. The gain against other tokenizers is the real one: GLM 5.3 is 3.93, DeepSeek V4 is 3.72, Kimi K3 is 3.28. A German document that costs 1,000 Kolibri tokens costs roughly 1,490 Kimi K3 tokens at those rates.

English web is a different story. Kolibri is 4.58 bytes per token. Plain BPE is 4.59. GPT-5 is 4.67, which is better. "Best German compression in this table" is accurate. "Better English compression than the frontier tokenizers" is not.

The word pictures are where UniBPE shows up. Kolibri splits *Bundessozialgerichtes* into Bundes / sozial / gericht / es. GPT-5, Qwen3.8, and Gemini split the middle into ess / oz / ial. *Protokolldaten* is Protokoll / daten for Kolibri and a five-piece cut for GPT-5. Cleaner morphemes are the claim. A large bytes-per-token win over their own BPE is not.

Grounding: it is trained to abstain #

Aleph Alpha's regulated-customer pitch is a model that says it does not know. Training includes "I don't know" targets plus their Merlin-Arthur procedure: one player adds evidence, one strips it, and the model has to answer only when the context still supports an answer. The paper is arXiv:2512.11614.

Kolibri is much harder to bait than Origin. It is not the most reluctant model in the set: Qwen3.6's non-hallucination rate is 56.7 against Kolibri's 44.0. The M/A score, which Aleph Alpha describes as a lower bound on how much of an answer came from the document, is where Kolibri leads (0.23 vs 0.12).

Sector scores, on their own suites #

Public benchmarks miss German public administration, aerospace, and factory language, so Aleph Alpha built private suites and synthetic training environments. They say no customer data went into training. These are still vendor evals of vendor tasks.

Kolibri leads four of the seven. Nemotron 3 Super leads cleaned MuSiQue and the German public-sector suite (78.0 vs 75.0). Aerospace is a tie with Qwen3.6, 58.9 against 59.0. A 99 on the automotive-supplier suite is a vendor number on a vendor task. Treat it as a clue about the training target, not as a score you can quote to a customer.

Run Kolibri with vLLM today #

This is Aleph Alpha's command. It serves weights you download. It does not call api.routerplex.com.

Install the plugin, which also installs the vLLM build it supports, or use the container ghcr.io/aleph-alpha/aleph-alpha-inference:

pip install "aleph-alpha-inference>=1.0"
 
vllm serve Aleph-Alpha/Kolibri-1 --kv-cache-dtype fp8 \
  --reasoning-parser kolibri1 \
  --tool-call-parser kolibri1 \
  --enable-auto-tool-choice

Contexts past 262,144 tokens need an override. Aleph Alpha's line is:

vllm serve Aleph-Alpha/Kolibri-1 --kv-cache-dtype fp8 \
  --reasoning-parser kolibri1 \
  --tool-call-parser kolibri1 \
  --enable-auto-tool-choice \
  --max-model-len 1048576 \
  --hf-overrides '{"max_position_embeddings": 1048576}'

Recommended sampling is temperature=1.0, top_p=0.97, top_k=128. Pass reasoning effort as none, low, medium, or high through the chat template. Leaving it unset selects high.

The FP8 checkpoint is about 78 GB. The model card's minimum is 2× A100 80 GB, 2× H100 SXM5, 1× H200, 1× B200, or 1× B300. The recommended set drops the A100s: 2× H100 SXM5, 2× H200, 1× B200, or 1× B300. That is the whole model in memory. Active parameters only reduce the math, not the footprint.

Not on RouterPlex, and what is #

RouterPlex: Kolibri will be on this catalog once it is listed. Hugging Face availability is not that bar. There is no model ID to paste into Cursor, Claude Code, or the OpenAI SDK, and there is no price to quote.

When a row exists it will be vendor list price, 0% markup, prepaid, with a hard per-key budget. The request stops at $0. The OpenAI-compatible chat API is the contract those rows use. Aleph Alpha has not published a list price, so this page will not invent one.

The comparison models that already have a hosted route are live here:

• Qwen3.8-Max at $2 / $6 per 1M, the dense model that leads Aleph Alpha's overall row. See Qwen3.8-Max API pricing.
• Qwen3.8 Flash at $0.15 / $0.47 per 1M when you want the cheap sibling.
• GLM-5.3 at $1.40 / $4.40 per 1M. The tokenizer table above is one reason German text can cost more on that route than on Kolibri's own tokenizer.

curl https://api.routerplex.com/v1/chat/completions \
  -H "Authorization: Bearer $ROUTERPLEX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3.8-max",
    "messages": [
      {"role": "user", "content": "Fasse die Risiken dieses Migrationsplans auf Deutsch zusammen."}
    ]
  }'

That ID is Qwen3.8-Max. It is not Kolibri. Another open-weight model waiting on a catalog row is Aikido Altar-1, a security prune of GLM-5.3.

Weights this week, a route when it is listed #

Download Aleph-Alpha/Kolibri-1 and serve it with the vLLM command above if you need the open-weight model this week. Create a RouterPlex account if you want the prepaid key ready when Kolibri is listed, and use the models already on the marketplace until then. This page will name the RouterPlex model ID on the day the catalog row exists.