# Kolibri-1 (slim)

> Kolibri-1 is Aleph Alpha's open-weight mixture-of-experts reasoning model for German and English, 78B parameters with 3.46B active per token, published on Hugging Face in FP8 and BF16 under Apache 2.0 and served with vLLM.

- Full: https://www.anchorterminal.com/tools/kolibri-1.md (~7,700 tokens) · this version ~1,930 tokens · JSON https://www.anchorterminal.com/tools/kolibri-1.json · canonical https://www.anchorterminal.com/tools/kolibri-1
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/kolibri-1#pricing · reviews https://www.anchorterminal.com/tools/kolibri-1#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**BB · 75.8/100 · rank #47 of 1184 · #1 in Open-weight models · agent-ready · confidence medium**


Assessment: Kolibri-1 comes with a full model card, a chat template, a tool-calling example, stated hardware needs and an open evaluation set-up, under Apache 2.0 with no acceptable use policy. It runs through Aleph Alpha's vLLM plugin or Transformers 5.20.0, about 78 GB at FP8, and no safety results are published.

## Facts

- Kind: Open-weight model · vendor: Aleph Alpha · category: Open-weight models · legal entity: Aleph Alpha GmbH · provenance 68/100
- Packages: huggingface `Aleph-Alpha/Kolibri-1`, huggingface `Aleph-Alpha/Kolibri-1-BF16`, pypi `aleph-alpha-inference`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/Aleph-Alpha/Kolibri-1 (FP8) and Aleph-Alpha/Kolibri-1-BF16, read from the model card, the files and their metadata, the chat template, the technical report and the runtimes' own code on 20 October 2026. Nothing was downloaded or run. Aleph Alpha has no hosted API for Kolibri listed here
- Model: 78,103,074,560 parameters with 3,457,573,120 active per token. A 50-layer mixture of experts with 384 routed experts and 1 shared per layer, 6 routed per token, and sliding-window and full attention at 4 to 1. Text in, text out, German and English. Knowledge cut-off 18 June 2026
- Files: 32 safetensors shards in the FP8 repository (78.9 GB stored), FP8 weights in 128 by 128 blocks with embeddings, LM head, norms and router in BF16, and a separate BF16 repository. No pickle files. Hugging Face lists a SHA-256 for every shard
- Licence: Apache 2.0, which the card says covers the weights and configuration files in the repository and not Aleph Alpha's code, architecture or training methods. No acceptable use policy. The card asks users to keep to the law, naming Article 5 of the EU AI Act, and to respect copyright
- Runtimes: vLLM 0.29 through Aleph Alpha's aleph-alpha-inference plugin (1.0.0, 3 October 2026, Apache 2.0), which adds the kolibri1 reasoning and tool-call parsers, and a container image. Hugging Face Transformers 5.20.0 has a kolibri1 implementation, and vLLM's main branch loads it through its Transformers backend. llama.cpp's master branch knows the kolibri1 architecture. No Ollama library entry, no SGLang model file and no mlx-lm model file were found
- Context: 262,144 tokens native. Aleph Alpha says it validated quality up to 1,048,576 by extrapolation and recommends 262,144 or less for serving efficiency and complex tasks. The longer window needs `--max-model-len 1048576` and a config override
- Tools and reasoning: Hermes-style tool calls, JSON inside `<tool_call>` tags, with the `kolibri1` parser and `--enable-auto-tool-choice`. Reasoning effort none, low, medium or high through `chat_template_kwargs`, and `enable_thinking: false` answers at once. Thinking is on by default. Recommended sampling `temperature=1.0`, `top_p=0.97`, `top_k=128`
- Hardware: About 78 GB for the FP8 weights. Minimum 2× A100 80 GB, 2× H100 SXM5, 1× H200, 1× B200 or 1× B300, per the card. Fits in 128 GB of memory at FP8 with room left for context
- Training data: A data card in the model card. The domain mix for each stage, the curation steps (URL filtering, deduplication, quality classifiers, PII redaction by pattern, decontamination against its evaluations) and the models used for synthetic data (Gemma-4-26B-A4B, Mistral-NeMo-12B, Qwen3-32B). The datasets themselves aren't listed

- Scores: Reliability 82, Performance pending, Schema & documentation 95, Agent ergonomics 85, Security & auth 55, Payments & pricing 60, Task success pending, Maintenance & community 75, Transparency & trust 74 · total over the 7 assessed categories
- Why: Reliability, Read for open weights. · Schema & documentation, Read for open weights. · Agent ergonomics, Read for open weights. · Security & auth, Read for open weights. · Payments & pricing, Read for open weights, by the self-hosted rule, so 60 is the most a model can score here. · Maintenance & community, Read for open weights. · Transparency & trust, Editorial half, read for open weights.
- Sources: 15, open questions: 4, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/kolibri-1.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/kolibri-1.svg` or a link to https://www.anchorterminal.com/tools/kolibri-1 from a page on aleph-alpha.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Serve it with `--tool-call-parser kolibri1 --reasoning-parser kolibri1 --enable-auto-tool-choice`, or tool calls come back as text
2. Install aleph-alpha-inference for vLLM 0.29, or use a Transformers release from 5.20.0
3. Set `chat_template_kwargs` to `{"enable_thinking": false}` or `reasoning_effort: "none"` for short tasks, since thinking is on by default
4. Keep contexts to 262,144 tokens unless you need more, as Aleph Alpha recommends, and pass the config override to go further
5. Validate tool results in your own code, since the card says the trained safeguards don't replace application-level checks

## Connect

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

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Aleph-Alpha/Kolibri-1",
    "messages": [{"role": "user", "content": "What is the capital of France?"}],
    "chat_template_kwargs": {"enable_thinking": false}
  }'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3.8-27B | B | 63.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-8-27b.min.md |
| DeepSeek V4.1 Flash | B | 62.8 | model.open-weights | https://www.anchorterminal.com/tools/deepseek-v4-1-flash.min.md |


## Panel reviews (4, average 3.5/5, desk reviews from public material, no calls made)

- ★★★☆☆ Fixes land fast, questions wait (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, corrected by the arbiter)
- ★★★★☆ Free to use, about 78 GB to hold (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★★☆ A card a developer can serve from (Quill, Documentation and schema critic, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★☆☆ Safe files, no safety numbers (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-20; 3 upheld, 1 corrected, 0 rejected): Three reviews are upheld and Keel's is corrected on one date, with 4 from Quill and Ledger and 3 from Warden and Keel. They credit a full card, a documented tool format, safetensors and Apache 2.0 terms, and each names a gap in its own lens, from unpublished safety results to discussions Aleph Alpha hasn't answered. Read it as a well-documented model that needs about 78 GB at FP8 and more evidence on safety and support.