Kolibri-1 in a Real Agent Test: 78B MoE With 3.46B Active Params and 1M Context Falls Short?
WolframRvnwlf · x · 2026-10-08
Aleph Alpha released Kolibri-1, an Apache-2.0 open-weight MoE built natively for German and English: 78.1B total parameters with only 3.46B active per token, controllable reasoning, tool calling, and up to 1M token context — an attractive package for sovereign, locally hosted agents.
The author went beyond benchmarks with a demanding persistent-agent test, self-hosting the pinned checkpoint on an H200 via Aleph Alpha's official inference plugin and vLLM 0.29, using FP8 weights and KV cache, official chat template, and documented sampling defaults. The "big bird with small wings" framing hints at a gap between benchmark scores and real agent behavior.
More from coding & agent
- Jev 0.4.0 brings semantic ranking and routing to PowerShell pipelines — dfinke · 2026-10-08
- Five AI agents handle a mid-project requirement change in human-agent platform Teamily — cneuralnetwork · 2026-10-08
- Watching my AI agent do my job while I just say 'looks good' and 'continue' — tekbog · 2026-10-08
- AAA graphics in a browser tab: Opus 5.5 builds a custom three.js water and lighting engine — Daniel_Farinax · 2026-10-08
- SpaceXAI donates $1.5M in Grok tokens to DHH's open-source Omarchy Linux — elonmusk · 2026-10-08
- Resume failed PydanticAI agent workflows without repeating tool calls via AgentRunResult — KhuyenTran16 · 2026-10-08