Primary use stated or inferred cautiously from official documentation.
AA
ALEPH ALPHA / Aleph-Alpha/Kolibri-1
Kolibri 1
Aleph Alpha's text and code model, officially released on October 3, 2026, focused on German and English. Its MoE architecture has 78.1 billion total parameters and activates about 3.46 billion per token. It supports configurable reasoning and tool calling. FP8 and BF16 weights are released under Apache 2.0; they are two precisions of the same family, not two independent models.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | Aleph-Alpha/Kolibri-1 |
|---|---|
| Model type | Text and reasoning · Code |
| Access model | Open weights |
| License | Apache 2.0 |
| Deployment | Self-hosted with vLLM |
| Release | 03 OCT 2026 |
| Knowledge cutoff | 18 JUN 2026 |
| Entrance | Text |
| Exit | Text |
| Context window | 1,048,576 tokens; recommended ≤262,144 |
| maximum output | Not published |
| Reasoning | none, low, medium, high |
| Published tools | Function calling · Structured Outings |
| Structured Outings | Supported in self-hosted deployments |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · vLLM |
Bilingual RAG, structured extraction, coding and internal-document analysis with human review. Native trained context is 262,144 tokens; the provider validates extension to 1,048,576 but recommends staying within 262,144 for complex tasks and serving efficiency. A Spanish profile does not mean the model is optimized for Spanish.
Weights occupy approximately 78 GB in FP8 and 156 GB in BF16, excluding full runtime memory and the KV cache. A small active-parameter count reduces compute, not the memory required to load all experts. Deployment requires the aleph-alpha-inference plugin for vLLM; check hardware, concurrency, errors and permissions before automating actions.
03 / CAPABILITIES
What it can do
Orientation
Bilingual RAG, structured extraction, coding and internal-document analysis with human review. Native trained context is 262,144 tokens; the provider validates extension to 1,048,576 but recommends staying within 262,144 for complex tasks and serving efficiency. A Spanish profile does not mean the model is optimized for Spanish.
Context
1,048,576 tokens; recommended ≤262,144 · Not published
Tools and integration
Function calling · Structured outputs
Access
Hugging Face · FP8 and BF16
Input modalities
Text
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
No official public per-token rate or commercially hosted endpoint has been verified. Downloading permissively licensed weights does not make inference free: budget for GPUs, memory, storage, operations and power. Aleph-Alpha/Kolibri-1 is the repository and self-hosted server identifier in the official example, not a promise of availability on a third-party API.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Figures published by Aleph Alpha, not independent Inferama measurements. We retain each test's language and version: a German translation of GPQA or AIME is not equivalent to the English dataset. Agentic results depend on the harness, tools and reasoning budget. Do not transfer FP8 scores to another quantization or mistake a validated million-token context for equal quality on every task.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| GPQA Diamond (EN) | 84,3 % | Score | View source ↗ |
| GPQA Diamond (DE) | 81,3 % | Score | View source ↗ |
| AIME 2025 (EN) | 96,9 % | Score | View source ↗ |
| AIME 2025 (DE) | 87,5 % | Score | View source ↗ |
| TerminalBench 2.1 | 27,7 % | Score | View source ↗ |
| Tau2-Bench (Telecom) | 94,7 % | Score | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Kolibri 1 · FP8
Aleph-Alpha/Kolibri-1: FP8 weights and FP8 KV cache in the published evaluation.
Kolibri 1 · BF16
Aleph-Alpha/Kolibri-1-BF16: BF16 precision variant of the same family.
ANALYSIS
Related analysis
07 / SOURCES