Language modelOpen weights

Kolibri 1

Aleph Alpha

Released: October 3, 2026
Data date: October 7, 2026

Input: Text
Output: Text

Kolibri 1 is Aleph Alpha's German-English MoE model with 78.1 billion total parameters and 3.46 billion active parameters. The provider lists 1,048,576 tokens as its validated context, but recommends no more than 262,144 tokens for serving efficiency and complex tasks. The weights use Apache 2.0.

Specifications and access

Specification	Value and source
Provider model ID	Aleph-Alpha/Kolibri-1Source
Access	Locally deployable open weightsSource
Input modalities	TextSource
Output modalities	TextSource
License	Apache 2.0Source
Context recommendation	Aleph Alpha recommends no more than 262,144 tokens for serving efficiency and complex tasks.Source
Reasoning and tools	Explicit reasoning mode and tool callingSource

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

Overall (EN) (Show measurement, test conditions, and source)

Source value: 75.5
Score: 75.5
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: aggregate
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

Overall (DE) (Show measurement, test conditions, and source)

Source value: 70.8
Score: 70.8
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: aggregate
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

Knowledge average (EN) (Show measurement, test conditions, and source)

Source value: 50.1
Score: 50.1
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: aggregate
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

Knowledge average (DE) (Show measurement, test conditions, and source)

Source value: 57.6
Score: 57.6
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: aggregate
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

GPQA Diamond (EN) (Show measurement, test conditions, and source)

Source value: 84.3
Score: 84.3
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: knowledge
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

GPQA Diamond (DE) (Show measurement, test conditions, and source)

Source value: 81.3
Score: 81.3
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: knowledge
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

Humanity's Last Exam (EN) (Show measurement, test conditions, and source)

Source value: 21.5
Score: 21.5
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: knowledge
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

Humanity's Last Exam (DE) (Show measurement, test conditions, and source)

Source value: 15.9
Score: 15.9
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: knowledge
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

AA-Omniscience Accuracy (public set) (Show measurement, test conditions, and source)

Source value: 14.8
Score: 14.8
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: knowledge
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

AA-Omniscience Index (public set) (Show measurement, test conditions, and source)

Source value: -32.8
Score: -32.8
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: knowledge
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

MMLU-Pro CoT (EN) (Show measurement, test conditions, and source)

Source value: 80
Score: 80
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: knowledge
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

MMLU-ProX CoT (DE) (Show measurement, test conditions, and source)

Source value: 75.5
Score: 75.5
Metric: reported benchmark score; source does not state unit
Unit: No unit provided
Category: knowledge
Direction: Higher is better
Source type: vendor-reported
Evaluator: Aleph Alpha Kolibri 1 model card
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Aleph Alpha internal evaluation table, exact post-training column labelled Kolibri only. The model card documents float8_e4m3fn block weights, dynamically quantized activations and an FP8 KV cache; the results use high reasoning effort and the model-specific context/sampling settings. The card says eval-framework is used for most benchmarks and Harbor for TerminalBench 2.1 and SWE-Bench. The table states no score unit or sample counts. It omits results when tool use is unsupported or context is exceeded; overlength long-context prompts score 0.
Context: Vendor-reported scores for the model card's documented FP8 checkpoint. Only the exact post-training Kolibri column is included; Kolibri Origin, Kolibri Base, BF16 sibling weights and the separate pre-training table are excluded. Published category means and overall rows are source-defined aggregates, not a shared independent cohort.

Source Aleph Alpha Kolibri 1 model card

Page 1 of 7

Sources and data date

Type	Evidence and data date
Input	Documented modalities. Retrieved October 7, 2026. Source: Aleph Alpha Kolibri 1 model card · Aleph Alpha Kolibri 1 model card (retrieved October 7, 2026) · Aleph Alpha Kolibri 1 model card · Editorial description reviewed October 7, 2026
Output	Documented modalities. Retrieved October 7, 2026. Source: Aleph Alpha Kolibri 1 model card · Aleph Alpha Kolibri 1 model card (retrieved October 7, 2026) · Aleph Alpha Kolibri 1 model card · Editorial description reviewed October 7, 2026
Research status	Research date October 7, 2026. 2 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not found in the inspected sources	independent-evaluation Original wording No independent evaluation of the exact Kolibri 1 post-training checkpoint was verified in this pass.
Additional source	MediSync original Kolibri-1 FP8 evaluation (retrieved October 7, 2026)