Aleph Alpha opens Kolibri, an Apache-2.0 78B MoE with 3.46B active
384 experts fire a thin slice per token, so it runs on one node — 262K context, tool calls, ~96% AIME. Plus a 2.7x open MoE trainer and agent routing.
Copy markdown
A reasoning MoE you can actually afford to run
Kolibri-1 is 78B total but fires only 3.46B params per token (384 experts, 6 routed), so it serves on a single node with `vllm serve Aleph-Alpha/Kolibri-1`. Apache-2.0, 262K native context (extends to 1M), tool calling, and a low/medium/high reasoning switch — it posts ~96% on AIME and 89% on code in English, no API dependency.
Train your own large MoE 2.7x faster
Ai2 shipped Olmo-core 3, a fully open training stack it says delivers 2.7x the throughput of Megatron on large mixture-of-experts runs, tested up to a 1.2T-parameter setup. If you pretrain or fine-tune MoEs, it's the open alternative to closed infra — released Oct 2.
Route each coding task to the cheapest model that nails it
An open-source model router for coding agents hit the Show HN front page, claiming Astra-level results by sending each task to the best-fit model instead of paying top-tier rates for everything. It builds on Agent-as-a-Router research showing routing across frontier LLMs beats any single model on cost-adjusted coding benchmarks.