18 September 2026
How to read an agent leaderboard honestly
Error terms, session counts and why a 4.8% net improvement is not a rounding accident.
ReleaseOpen weights
Yiğido 1.0 is out: open weights, published evals
The first Turkish foundation model to ship weights, a datasheet and a leaderboard entry at the same time.
16 September 2026 · 4 min read · Vincent Loveqcn · Yiğido Lab, İstanbul
Yiğido 1.0 is live. Two models, one recipe, everything published:
- Yiğido 1.0 (Max) — 32B MoE (4B active), 128K context, 1.7T pre-training tokens.
- Yiğido 1.0 Mini — the same mixture, distilled to run on a single 24 GB card.
What “open” means here
Weights, tokenizer, data datasheet and the evaluation harness are in the release. No research-only clause, no contact form. If you run a bank in Istanbul with strict data-residency rules, you can host Yiğido inside your own network.
What the board says
On the Agent Arena “Overall” snapshot of 15 September 2026, Yiğido 1.0 (Max) enters at rank 12 with +4.80% net improvement, 8.60% confirmed success and 2.10% steerability. That places it above several closed frontier models on confirmed success, and behind the top Anthropic and OpenAI runs.
Rank 12 is the honest number, and it is the number we will keep publishing — including the ± error terms, including the runs where we lose.
What is next
The 1.1 branch targets three things: longer Turkish documents, better tool-call reliability, and a cheaper Mini that still fits your laptop.