Wait Which Model?
← Back to directory
Aleph Alpha

Kolibri-1

Released Oct 3, 2026 · knowledge cutoff 2026-06

Status
Superseded
Location
Germany
Type
LLM
Modality
Text
Context window
1M
Max output
—
Speed
—
Price ($/MTok in / out)
— / —
Cost per task
—
Open weights
Yes
Benchmarks6 of 10 reported

MMLU-Proretired

80%

GPQA Diamond

84.3%

SWE-bench Verified

66.4%

Terminal-Bench 2.1

27.7%

AIME (latest)retired

96.9%

Humanity's Last Exam

21.5%

SWE-bench Pro

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Trained to say it doesn't know when the answer isn't in the supplied context — Aleph Alpha built training examples where abstaining is the only correct reply, aimed at document work where a wrong answer costs more than none
  • Bilingual tokenizer built around German compound words, so German documents split into fewer tokens and leave more of the window free than on English-centric models
  • Reasoning effort is selectable from none to high, so quick lookups and harder problems can share one deployment
  • Apache 2.0 weights sized for on-premise use — a single B200 or a pair of H100s holds the FP8 checkpoint, so sensitive documents never have to leave the building

Weaknesses

  • Not a drop-in for stock vLLM — reasoning and tool calling need Aleph Alpha's own inference plugin, an extra dependency in any self-hosted stack
  • No hosted endpoint at launch — trying it at all means pulling roughly 78 GB of weights onto data-centre GPUs
  • Too new to characterise — there are no independent evaluations or hands-on reports yet, so its behaviour rests on Aleph Alpha's own claims

Aleph Alpha's English-German reasoning model, a sparse MoE with 78.1B total and 3.46B active parameters (384 routed experts per layer, 6 active), developed in Germany and trained on infrastructure in Germany and Finland. Announced and released on Hugging Face on 2026-10-03; the blog brands it simply "Kolibri" and the repo is Kolibri-1 (FP8, with a BF16 variant). Context is 262,144 tokens natively, extended to 1,048,576 by extrapolation, which the blog advertises as "up to 1M". Every score is Aleph Alpha's own, from the model card and launch post at high reasoning effort: GPQA Diamond 84.3 (81.3 in German), AIME 2025 96.9 (AIME 2026 96.0), MMLU-Pro CoT 80.0, SWE-bench Verified 66.4 (scaffold unstated), HLE 21.5 (tool use unstated), Terminal-Bench 2.1 27.7, plus tau2-Bench Telecom 94.7, BFCL v4 61.4 and BrowseComp 29.4. Availability is self-host: there is no Aleph Alpha API SKU and no OpenRouter listing, and serving needs the aleph-alpha-inference vLLM plugin on at least two A100 80GB GPUs. No pricing, max output or Artificial Analysis entry as of 2026-10-05. Knowledge cutoff is 18 June 2026 for both languages. predecessorId is null: the launch post presents Kolibri Origin (30B-A3B, finished June 2026) as the model that validated the training pipeline, not as one Kolibri replaces, and Origin is not on this site.

For developers

API model strings

Not researched

Licence

Apache License 2.0 · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor

News