Inkling 2 Small Launch - ECTSum Fine-tuning Step 200: leaderboard
Public earnings-call financial-document task ECTSum. All models trained on Tinker using Tinker Cookbook Async RL / Stable Reinforce, with a learning-rate sweep per model, for 200 steps. Inkling models are means of three seeds; comparators single runs. Admit the declared training endpoint, retaining the full native trajectory as context. Estimated dollar costs and intermediate training checkpoints are not separate capability boards. Trained checkpoints are not stock base-model weights.
Metric: Held-out rubric reward. Source: thinkingmachines.ai. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Inkling 2 Small (Benchmark Sweep Run: source=Inkling2SmallECTSum; Stable Reinforce Async RL; per-model LR sweep; step=200; 3-seed mean) | 0.94 | |
| 2 | DeepSeek V4.1 Flash 1017 (Benchmark Sweep Run: source=Inkling2SmallECTSum; Stable Reinforce Async RL; per-model LR sweep; step=200; single run) | 0.85 | |
| 3 | GLM 5.3 Flash (Benchmark Sweep Run: source=Inkling2SmallECTSum; Stable Reinforce Async RL; per-model LR sweep; step=200; single run) | 0.83 | |
| 4 | Inkling (Benchmark Sweep Run: source=Inkling2SmallECTSum; Stable Reinforce Async RL; per-model LR sweep; step=200; 3-seed mean) | 0.82 | |
| 5 | Inkling-Small (Benchmark Sweep Run: source=Inkling2SmallECTSum; Stable Reinforce Async RL; per-model LR sweep; step=200; 3-seed mean) | 0.79 | |
| 6 | Qwen 3.5 397B (Benchmark Sweep Run: source=Inkling2SmallECTSum; Stable Reinforce Async RL; per-model LR sweep; step=200; single run) | 0.73 |
Interactive version: theaggregate.ai/benchmark?slug=inkling-2-small-launch-ectsum-fine-tuning-step-200 · How It Works · Data refreshed daily, snapshot 2026-11-06.