Kolibri-1

Spec

Rows with no slot in this schema, all from the same source:

Release Date

2026-10-03, announced and released the same day. Weights were published to Hugging Face on that date.

Benchmarks

As published by Aleph Alpha (source). The comparison set is Aleph Alpha's own:

No frontier closed model appears in the table, and no figure for any Claude, GPT, Gemini or Qwen model is stated, so this page carries no comparison against the frontier.

Use Cases

The announcement positions the model for sovereign and on-premises deployment — Apache 2.0 weights, published in full, for operators who need the model to run on their own hardware. The bilingual training mix (4.3 trillion German tokens, 21.3% of a 20 trillion token three-stage pre-training run) targets German-language work specifically, which is the gap Aleph Alpha's own earlier writing argues for.

The extended 1,048,576-token context is stated as an extension of a 16,384-token native window, so long-context use rests on that extrapolation rather than on native training length.

No deployment guidance, hardware requirement or serving configuration is stated in the first-party post. See ## Conflicting Reports.

Compared To

• Nemotron 3 Super 120B and Mistral Small 4 119B — the two models Aleph Alpha's own table compares against; see ## Benchmarks above. A Mistral Small 4 release is held on this wiki under Mistral AI.
• Qwen 3.8 27B — the other current open-weight model this wiki tracks in the locally-runnable class. No head-to-head figure exists in anything read; the two are not compared here.
• Sparsity as a design axis is the subject of a scaling result captured the same day — see Test-Time Compute (Inference-Time Compute Scaling) on Scaling Laws for Looped Mixture of Experts, which reports ~3× active-parameter efficiency from sparsity as a general finding. That paper does not evaluate Kolibri-1, and the two are linked as the same question, not as evidence for each other.

Sources

• Kolibri Has Landed: A Sovereign Open-Weight Model — Aleph Alpha, 2026-10-03, read first-party 2026-10-04

Conflicting Reports

Native context window: 16,384 or 262,144 tokens. The first-party announcement states 16,384 native, extended to 1,048,576. Secondary coverage published 2026-10-03/04 — MarkTechPost, testingcatalog, alphasignal and daily.dev — states 262,144 native, extended to 1,048,576 (source).

The first-party figure is the one in ## Spec, per the CLAUDE.md priority rule (official announcement > third-party reporting). The disagreement is recorded rather than resolved because it is not a rounding difference — a 16× gap in native length changes what the extrapolation is doing — and because the technical report that would settle it did not parse.

Figures only secondary coverage states, and which are therefore not adopted onto this page: an FP8 checkpoint of about 78 GB, runnable on a single B200/B300/H200 or 2×H100 SXM5 through vLLM, and a 128,000-token vocabulary. The first-party post states no file size, no hardware and no vocabulary size.