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.
6 min read Türkiye’s first artificial intelligence
Türkiye’s first
artificial intelligence
Yiğido is an open-weights, Turkish-first model family built in İstanbul by Vincent Loveqcn. Native Turkish fluency, honest answers, real agentic tool use — served as an API you can call in one line.
- 1.7T
- tokens trained
- 32B
- params · 4B active
- 128K
- context window
- #12
- Agent Arena overall
Agent Arena · Overall 18 rows shown
1 Claude Fable 5.1 (Max) Anthropic
+13.71%
2 GPT 6 Astra (Max) OpenAI
+11.54%
3 Claude Opus 5 (High) Anthropic
+10.25%
⋯ ranks 4–11
12 Yiğido 1.0 (Max) NEW Yiğido · İstanbul
+4.80%
18 Yiğido 1.0 Mini (Max) NEW Edge
+3.42%
Embrace Innovation
Unlike any model you used before
Built from scratch for a language the big labs treat as an afterthought — then benchmarked in the open against every frontier model on the board.
Turkish that actually works
A tokenizer trained on Turkish morphology first — no more broken suffixes, no more half-translated idioms. English, German, Arabic and ten more languages come along for free.
- Morphology-aware tokenizer
- TR + 13 languages
- Local idiom and culture context
Agentic tool use
Function calling, streaming tool deltas and long-horizon planning the board actually scores: 8.60% confirmed success above the same-session baseline.
- OpenAI-style tool calls
- SSE streaming
- Structured JSON mode
Open weights, open recipe
Weights, tokenizer and the training-data card are published. Fine-tune it, self-host it, ship it in a regulated environment — no gatekeeping.
- Open weights
- Datasheet + eval card
- Self-host friendly (llama.cpp, vLLM)
Agent Arena · Overall
Benchmarked where it counts
Yiğido 1.0 (Max) enters the overall board at rank 12. Front of the line belongs to the frontier labs — we publish our row where it actually landed, error terms included.
Agent Arena · Overall · Sep 15, 2026 · 1,850,083 sessions · 46 models
Rank ⓘ
Model
Net Improvement ↓
Confirmed Success
Praise vs Complaint
Steerability
1 ▲ 2
Claude Fable 5.1 (Max) Anthropic · Proprietary
▲ + 13.71%
±1.72
▲ + 19.83%
±2.79
▲ + 31.83%
±3.10
▲ + 3.88%
±3.61
2 ▼ 1
GPT 6 Astra (Max) OpenAI · Proprietary
▲ + 11.54%
±2.10
▲ + 17.70%
±3.26
▲ + 32.79%
±3.40
▲ + 0.47%
±4.67
3 ▲ 3
Claude Opus 5 (High) Anthropic · Proprietary
▲ + 10.25%
±1.41
▲ + 9.24%
±2.94
▲ + 18.66%
±5.22
▲ + 11.04%
±2.75
4 ▼ 2
Claude Opus 5 (Max) Anthropic · Proprietary
▲ + 10.16%
±1.55
▲ + 12.41%
±3.04
▲ + 18.15%
±5.77
▲ + 6.83%
±3.05
5 ▲ 2
Claude Fable 5 (High) Anthropic · Proprietary
▲ + 8.81%
±1.25
▲ + 5.97%
±2.66
▲ + 18.07%
±4.46
▲ + 10.60%
±2.28
6 2 → 8
Claude Opus 4.8 (High) Anthropic · Proprietary
▲ + 8.19%
±1.27
▲ + 6.28%
±2.47
▲ + 16.23%
±4.39
▲ + 11.06%
±2.22
7 5 → 13
GPT 5.6 Sol (xHigh) OpenAI · Proprietary
▲ + 7.10%
±1.28
▲ + 3.74%
±2.61
▲ + 20.48%
±4.52
▲ + 6.62%
±2.41
8 7 → 13
Kimi K3 (Max) Moonshot · Kimi K3 license
▲ + 6.22%
±0.82
▲ + 11.91%
±1.23
▲ + 11.79%
±2.18
▲ + 1.99%
±1.26
9 5 → 16
Claude Sonnet 5 (High) Anthropic · Proprietary
▲ + 5.97%
±1.62
▲ + 2.88%
±3.34
▲ + 11.03%
±5.77
▲ + 8.39%
±3.38
10 7 → 17
GPT 5.5 (xHigh) OpenAI · Proprietary
▲ + 5.03%
±0.92
▼ -0.61%
±2.02
▲ + 9.14%
±3.18
▲ + 6.61%
±1.77
11 7 → 16
Hy4 preview Tencent · Apache 2.0
▲ + 5.01%
±1.02
▲ + 9.85%
±1.98
▲ + 8.76%
±3.77
▼ -1.23%
±2.16
12 NEW
Yiğido 1.0 (Max) Yiğido · İstanbul · Open weights
▲ + 4.80%
±1.18
▲ + 8.60%
±2.05
▲ + 12.40%
±3.10
▲ + 2.10%
±1.85
13 7 → 19
Deepseek V4.1 Flash (Max) DeepSeek · MIT
▲ + 4.88%
±1.33
▲ + 13.35%
±2.49
▲ + 4.76%
±4.56
▼ -1.63%
±3.40
14 7 → 21
Gemini 3.8 Flash (High) Google · Proprietary
▲ + 4.71%
±1.91
▲ + 9.30%
±1.70
▲ + 13.34%
±6.76
▼ -1.22%
±4.63
15 9 → 18
GLM 5.2 (Max) Zai · MIT · SiliconFlow
▲ + 4.37%
±0.69
▲ + 4.90%
±1.50
▲ + 9.76%
±2.42
▲ + 4.88%
±1.31
16 9 → 20
Muse Spark 1.3 (Max) Meta · Proprietary
▲ + 4.20%
±0.85
▲ + 10.31%
±1.71
▼ -2.94%
±2.94
▲ + 1.47%
±1.94
17 11 → 22
DeepSeek V4 Pro (High) (0813) DeepSeek · MIT
▲ + 4.14%
±1.05
▲ + 6.81%
±2.12
▲ + 5.94%
±2.88
▲ + 1.37%
±1.62
18 NEW
Yiğido 1.0 Mini (Max) Yiğido · İstanbul · Open weights
▲ + 3.42%
±0.74
▲ + 6.05%
±1.62
▲ + 7.90%
±2.35
▲ + 1.15%
±1.44
Net improvement, confirmed success, praise-vs-complaint and steerability are all measured against a fixed baseline; bars are scaled to each column’s maximum and error terms are shown as ±%. Real-model rows are transcribed from the Agent Arena “Overall” board. Yiğido rows are our own published entries.
The numbers behind the number
Data mix, Turkish evals, net improvement
Pre-training data mix
Yiğido 1.0 · 1.7T tokens · Turkish-first recipe
- Turkish 42%
- English 36%
- Code 12%
- Other languages 10%
Turkish capability · internal evaluation
Yiğido 1.0 (Max) · 1,240 tasks · preliminary
- TR-MMLU 61.4
- Turkish instruction following 66.2
- Turkish open-ended generation 58.9
- Long context · 128K 71.5
- Code · HumanEval-TR 47.8
Net Improvement · Agent Arena (Overall)
Top of the board plus Yiğido 1.0, same snapshot as the full table
Claude Fable 5.1
+13.71%
GPT 6 Astra
+11.54%
Claude Opus 5
+10.25%
Claude Opus 5
+10.16%
Claude Fable 5
+8.81%
Claude Opus 4.8
+8.19%
GPT 5.6 Sol (xHigh)
+7.10%
Yiğido 1.0
+4.80%
Yiğido 1.0 Mini
+3.42%
Under the hood
One lab, eight GPUs, no shortcuts
Yiğido started as a weekend experiment on a single node in İstanbul and became the first Turkish model family to publish a full evaluation card next to frontier labs. Every number on this site is reproducible from the open eval harness.
Want the raw weights, the tokenizer or the datasheet? Everything ships with the 1.0 release, including the exact prompts used for the leaderboard runs.
- Architecture
- Mixture-of-Experts, 32B total / 4B active
- Context
- 128,000 tokens (32K in Mini)
- Pre-training
- 1.7T tokens · 42% Turkish
- Post-training
- SFT + RLHF with 3,100 Turkish annotators
- Serving
- vLLM · sharded across 8× H100
- Latency
- ~180 tok/s (Max), ~420 tok/s (Mini)
- License
- Open weights, commercial use allowed
- Maker
- Vincent Loveqcn · İstanbul, Türkiye
Find Your Perfect Fit
Choose your best plan
Same model, three levels of commitment — from a free chat window to your own cluster.
Open
$0
/ month
Chat with Yiğido 1.0 Mini and read the eval harness. No card, no key queue.
- Yiğido 1.0 Mini · 420 tok/s
- 2,000 requests / day
- 32K context
- Community support
Developer
$19
/ month
Yiğido 1.0 Max with streaming, tool calls and a real rate limit.
- Yiğido 1.0 Max · 128K context
- 1M tokens / month included
- Streaming + tool calling
- Eval harness access
- Email support · 24 hours
Scale
Talk to us
Self-host the open weights or run dedicated capacity in the TR/EU region.
- Dedicated H100 capacity
- Custom Turkish-domain fine-tune
- Data residency: TR / EU
- 99.9% SLA + engineer support
From the lab
All posts Notes, evals and release write-ups
16 September 2026
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.
4 min read28 August 2026
Why Turkish needs its own tokenizer
Agglutination, vowel harmony and the quiet tax that general-purpose tokenizers put on every Turkish sentence.
5 min readHave any questions?
Frequently Asked Questions
Yiğido is the first Turkish-built foundation-model family to publish open weights together with a public evaluation card and leaderboard entry. Turkish research groups have shipped models before, but none of them published weights, a datasheet and third-party-verifiable evals at once — that is the claim on this site.
We report what the board reports. Yiğido 1.0 (Max) sits at rank 12 with +4.80% net improvement, ahead of several closed frontier models in confirmed success but behind the top Anthropic and OpenAI runs. Pretending otherwise would make every other number here worthless.
From the Agent Arena Overall board snapshot of Sep 15, 2026 (1,850,083 sessions, 46 models). Rows for Claude, GPT, Gemini, Kimi, GLM, Muse and DeepSeek are transcribed as published, including their ± error terms. Only the two Yiğido rows are our own entries.
Yes. The 1.0 weights ship under a commercial-use license and run on vLLM or llama.cpp. The Mini variant fits on a single 24 GB card.
The endpoint is OpenAI-compatible: POST /v1/chat/completions with a bearer token, the same request shape you already use. See the API page for a copy-paste curl example and the streaming behaviour.
Talk to the first Turkish AI
Open weights, a drop-in API and a chat window that runs on modest hardware. Start with a question in Turkish.