Unofficial, experimental GGUF conversion and quantization of Aleph Alpha Kolibri-1-BF16.
Requires the included Kolibri1 llama.cpp source patch. The unmodified llama.cpp revision used as the base for this port does not support this architecture. Compatibility with other releases, Ollama, or LM Studio has not been verified. This is an independent conversion; Aleph Alpha and the llama.cpp project have not endorsed it.
Deutsch: Dies ist eine unabhängige 3-Bit-GGUF-Version von Kolibri 1. Sie benötigt die hier dokumentierte experimentelle llama.cpp-Erweiterung. Das vollständige Modell wurde lokal auf einer RTX 3060 und einer Intel Arc Pro B60 gestartet; eine umfassende Bewertung der Antwortqualität steht aus.
| Property | Value |
|---|---|
| File | Kolibri-1-Q3_K_S.gguf |
| Size | Approximately 33.87 GB / 31.54 GiB; exact size in provenance.json |
| Format | GGUF v3, architecture kolibri1 |
| Scheme | Q3_K_S; mixed precision, approximately 3.47 bits per parameter overall |
| Tensor types | 501 Q3_K, 401 F32, 1 Q6_K |
| Router projections, norms, biases | F32 |
| Output matrix | Q6_K |
| Importance matrix | None |
| Further training | None |
“3-bit” describes the quantization scheme. Some tensors deliberately retain higher precision. All experts are included; active parameters per token do not represent the memory needed to store the model.
The original BF16 weights were converted with the included streaming converter and quantized with patched llama.cpp. On 2026-10-03, the publication copy received embedded license, source, and modification notices. Its entire quantized tensor payload is unchanged from the locally tested GGUF. See MODIFICATIONS.md, provenance.json, and SHA256SUMS.
The original model has a native context of 262,144 tokens. Aleph Alpha reports extended-context validation up to 1,048,576 tokens and recommends at most 262,144 for efficiency and complex tasks. Upstream model card
Initial validation used a 4,096-token server context. On 2026-10-24, additional local runs used a 32,768-token server context on RTX 3060 + Arc Pro B60. Those runs used short prompts; they do not validate a filled 32k context, the full upstream context, reasoning mode or tool calling. Larger contexts need additional KV-cache memory and runtime validation. See LOCAL-RUNTIME.md for the exact scope and limitations.
Build the pinned llama.cpp revision with the supplied patch using runtime-source/README.md. Use that resulting llama-server or llama-cli binary to load this file. Hardware selection and memory requirements depend on the machine; the model file alone needs approximately 31.54 GiB before runtime buffers and KV cache.
Initial server settings: 4,096-token context, one request slot, Q8_0 key/value cache, flash attention, Vulkan layer split 1:2 across NVIDIA RTX 3060 12 GiB and Intel Arc Pro B60 24 GiB. The AMD integrated GPU was excluded. The newer local launcher uses a 32,768-token context and a 0.85:2.15 layer split for Kolibri. See LOCAL-RUNTIME.md for configuration, recovery and local measurements; the existing runtime-source instructions remain the architecture-port build instructions.
Prior checks covered all 903 tensor names, shapes, offsets, types, and finite F32 router/norm values. The port was compared numerically against an independent full-precision reference on a small fixture, and tokenizer comparisons passed 169 cases. The full quantized model answered “Paris” to a capital question and “17 mal 23 ergibt 391.” to a multiplication question. Details are in validation.json.
These checks establish a limited functional result. They are not a language-quality benchmark, a safety evaluation, or evidence that upstream benchmark results transfer to this quantization. Quantization can reduce accuracy. No claim of unchanged capabilities or validated long-context behavior is made.
Original model provider: Aleph Alpha GmbH. Original model developer: Aleph Alpha Research GmbH. The original weights and published configurations are Apache-2.0 licensed. This quantized distribution includes the unchanged LICENSE, attribution in NOTICE, and modification details.
The model repository's license grant covers its published weights and configuration files. Other artifacts have their own rights and licenses. The included runtime source patch uses separately licensed Apache-2.0 inference reference code and MIT-licensed llama.cpp code; see runtime-source/THIRD_PARTY_NOTICES.txt.
The original model's intended uses, limitations, and responsible-use information remain relevant; consult its model card. Users remain responsible for complying with applicable law. Distribution is subject to the included licenses and their warranty disclaimers.
Provenance correction (2026-10-24): the originally recorded source revision 7a8f290e7858825c3cf5e4c447ba68345de9f1d3 is not resolvable on the upstream Hub. It is retained in the historical notices as an unverified record, not as a working reproducibility pin. The actual source revision used for conversion remains unresolved; no current upstream SHA is substituted as if it were the conversion source. This documentation update does not change the GGUF or its license.
Aleph-Alpha/Kolibri-1-BF16