Uncensored Kolibri-1: Aleph Alpha's 78B MoE Model Explained

Kolibri-1 at a glance

Developer

Aleph Alpha (Heidelberg, Germany)

Release date

October 3, 2026

Parameters

78B total, 3.46B active per token

Architecture

MoE Transformer, 50 layers, 384 experts

Context window

262,144 native, validated to 1,048,576

Languages

German and English

License

Apache 2.0 (weights and config)

Knowledge cutoff

June 18, 2026

Minimum hardware

2×A100 80GB, 2×H100 or 1×H200

Uncensored?

No, safety-aligned; community builds only

What is Kolibri-1?

Aleph Alpha Kolibri-1 is a mixture-of-experts (MoE) Transformer published by Aleph Alpha GmbH of Heidelberg, Germany, with Aleph Alpha Research as the developer. It has 78B total parameters (78,103,074,560) but activates only 3.46B per token, which the company rounds to about 3B. All 50 layers are MoE layers, each with one shared expert and six routed experts out of 384. Native context is 262,144 tokens, and Aleph Alpha says it has validated quality up to 1,048,576 tokens. The vocabulary of 128,000 tokens is tuned for German compounds. Only German and English are supported, the knowledge cutoff is June 18, 2026, and the model offers four reasoning-effort levels (none, low, medium, high) plus native tool calling.

Why people look for an uncensored Kolibri-1

The official model was trained with refusal and abstention behavior: Aleph Alpha's responsible-use terms forbid illegal uses and ask deployers to add their own guardrails, and the model is taught to say it does not know when evidence is missing. Writers and roleplayers who want frank adult fiction therefore look for an uncensored Kolibri-1. That only exists as unofficial community modifications such as abliterated weights, which need large hardware and come with no quality guarantees. Spicy AI does not host Aleph Alpha's model; its chat runs a different model with no content filters for consenting adults.

Kolibri-1 benchmarks vs Qwen3.6 and Nemotron 3 Super

All published scores are self-reported by Aleph Alpha; as of October 15, 2026 no independent reproduction had been found. In the launch blog, the model scores 96.9 on AIME 2025 versus 84.6 for Qwen3.6-35B-A3B and 91.7 for Nemotron 3 Super 120B-A12B, and 84.3 on GPQA Diamond versus 83.4 and 78.0. On the German versions it keeps a lead: 87.5 on AIME 2025 (DE) and 81.3 on GPQA Diamond (DE). Qwen3.6 wins on BFCL v4 (67.2 vs 61.4), LongBench Pro (70.8 vs 64.5) and τ²-bench telecom (99.1 vs 94.7). Aleph Alpha claims the model matches systems with up to four times more active parameters, naming Nemotron 3 Super.

Scores from Aleph Alpha's launch blog, scale 0–100. Not independently reproduced.

How Kolibri-1 was trained

Aleph Alpha documents the training run in unusual detail. Pre-training used about 20 trillion tokens at a 16,384-token sequence length, followed by 3.44 trillion mid-training tokens at 64k and roughly 201 billion long-context tokens at 256k, about 24 trillion in total. The job ran on 768 NVIDIA B200 GPUs in Germany and Finland, with pre-training taking 21 days. The model card puts the mix at about 62.5% English, 23.9% German and 13.6% code. Post-training combined a 268-billion-token SFT mix with reinforcement learning across more than 1.2 million tasks, and abstention data teaches the model to admit when context does not support an answer. Aleph Alpha says it designed the release around the EU AI Act, the GPAI Code of Practice and GDPR, and estimates training energy at about 950 MWh.

License: Apache 2.0 and its limits

The weights and configuration files are released under Apache 2.0, so you can download, fine-tune, redistribute and use them commercially, as long as you keep the license and attribution notices. The model card adds a boundary: the license covers only the files in the repository and does not extend to Aleph Alpha's underlying code, architecture, parameter settings or training methods, which remain the company's intellectual property. Self-hosting is free; Aleph Alpha has not published a per-token price for a hosted API, so enterprise access goes through its sales team.

Download from Hugging Face and GPU requirements

There are only two first-party repositories: Aleph-Alpha/Kolibri-1 with FP8 weights (about 79 GB) and Aleph-Alpha/Kolibri-1-BF16 (about 156 GB). FP8 needs roughly 78 GB of GPU memory. The documented minimum is 2×A100 80GB, 2×H100 SXM5, 1×H200, 1×B200 or 1×B300; Aleph Alpha says two H100 cards can serve 18 concurrent 256k-token requests. Serving goes through vLLM with the aleph-alpha-inference plugin: run vllm serve Aleph-Alpha/Kolibri-1 --kv-cache-dtype fp8 --reasoning-parser kolibri1 --tool-call-parser kolibri1 --enable-auto-tool-choice. To go past 262,144 tokens add --max-model-len 1048576 with a max_position_embeddings override. Recommended sampling is temperature 1.0, top_p 0.97 and top_k 128.

GGUF quants, llama.cpp and Ollama

Aleph Alpha ships no GGUF, ONNX or MLX files. The most downloaded conversion is Hob-forge/Kolibri-1-GGUF, whose author states it is independent and not endorsed by Aleph Alpha or llama.cpp. Sizes range from Q2_K at 28.6 GB through Q4_K_M at 47.5 GB to Q8_0 at 83.1 GB. As of its October 8 update, upstream llama.cpp did not support the kolibri1 architecture, so you need the repository's patch or the author's prebuilt binaries, and Ollama and LM Studio cannot load it yet. The author reports about 12.3 tokens per second for Q4_K_M on a Ryzen 7 7800X3D with 128 GB RAM, rising to about 18 with attention on a 12 GB GPU; these are self-reported figures, not benchmarks.

Sizes from the Hob-forge repository; usage notes are our own rough guidance.

Abliterated and uncensored community builds

Several community variants try to remove refusals. huihui-ai/Huihui-Kolibri-1-BF16-abliterated-GGUF strips the refusal direction from the weights; iSkye/Kolibri-1-heretic and Blackfrost-AI/KOLIBRI-1-DERISKED-BF16 follow similar ideas. None of them is published or tested by Aleph Alpha, nobody has published benchmarks showing how much reasoning quality they lose, and they need the same 30 to 160 GB of memory as the originals. Removing refusals also does not change the license or the law: content involving minors, real people without consent or other illegal material stays prohibited. For adult roleplay without that setup, an online uncensored chat is the practical route.

Our take: strengths and weak spots of Kolibri-1

Reading the model card and launch material side by side, Kolibri-1 looks less like a general chatbot and more like an infrastructure model for European organizations that must keep data on their own servers. Its best arguments are efficiency and German quality, not raw creativity. These are our editorial conclusions from the published facts, not test results of our own:

• Strength: only 3.46B active parameters means fast decoding for its size, which matters when many users share one GPU node.
• Strength: German is a first-class language, with a tokenizer tuned for long compound words and German benchmark versions reported.
• Strength: Apache 2.0 weights plus a published data summary make legal review easier than for many open models.
• Weak spot: you still need about 78 GB of GPU memory, so it is not a laptop model despite the small active count.
• Weak spot: other languages, including Chinese, are not officially supported or evaluated.
• Weak spot: tool-calling and long-context scores trail Qwen3.6 in Aleph Alpha's own tables.

Early community reception

Two weeks after launch there are no independent benchmark reproductions or long-term user reviews we could verify, so we do not quote any. What can be measured is activity on Hugging Face, which shows strong curiosity from people who want to run Kolibri-1 outside a data center. Download counts show interest, not quality:

• Hob-forge/Kolibri-1-GGUF had 42,241 downloads and 26 likes by October 15, 2026, the most of any conversion.
• huihui-ai's abliterated GGUF reached 4,551 downloads, a sign that many users want fewer refusals.
• Apple-silicon users published 3-bit and 4-bit MLX conversions, and others built NVFP4 and EAGLE3 drafter variants.
• A community Gradio Space appeared on October 4, but Aleph Alpha itself offers no public demo chat.

Not the Kolibri 1/2 textbook, Kolibri app or Colibri LLM

Kolibri is a crowded name, so check what a search result refers to. Kolibri 1/2 is a German primary-school textbook series, and Kolibri download pages usually point to Learning Equality's offline education app for Windows and Android. Colibri LLM, often spelled with a C, is an unrelated inference engine on GitHub. The model on this page is always Aleph Alpha Kolibri-1, hosted under the Aleph-Alpha organization on Hugging Face; its predecessor was called Kolibri Origin.

Prompt tips for Kolibri-1-style results

1. The model only officially supports German and English, so local runs work best in those languages.
2. Use reasoning_effort none or low for chatty roleplay; high effort adds long thinking you rarely need.
3. Keep context at or below 262,144 tokens unless you really need the 1M mode.
4. Treat abliterated builds as experiments and compare them with the original before relying on them.
5. State that every character is an adult (18+) and never base a scene on a real person.

Sources

• Kolibri-1 model card (Hugging Face)
• Kolibri-1 BF16 weights (Hugging Face)
• Kolibri Has Landed: launch blog (Aleph Alpha)
• Press release (Aleph Alpha)
• Hob-forge Kolibri-1 GGUF (third party)
• Huihui abliterated GGUF (third party)

More uncensored AI Chat models

• Uncensored ChatGPT
• Uncensored Claude
• Uncensored Gemini
• Uncensored Grok
• Uncensored DeepSeek
• Uncensored Llama
• Uncensored Mistral
• Uncensored Qwen
• Uncensored Character.AI
• Uncensored Kimi
• Uncensored Gemma