Indian STEM reasoning models — solver, table recognition, and multimodal QA. GPU inference runs serverless on Modal; the first call after idle wakes the GPU (~1–2 min), then it's fast.
Nalanda Qwen 7B (GRPO / RLVR-tuned) solves a physics, chemistry, or maths problem with a worked solution.
DrishtiTable (Qwen2.5-VL-7B + LoRA) converts a table image into clean HTML.
nalanda-image-vl answers a question about a scientific diagram or figure.
Note: this model builds on Meta's gated Llama-3.2-Vision. If access is still pending on the backend token, this tab will return an error until it's granted.
Image → HTML table recognition on 135 held-out Indian-textbook tables. Ranked by TEDS (the field-standard TSR metric). Our fine-tuned 7B model beats every zero-shot frontier model.
| Model | Method | TEDS | S-TEDS |
|---|---|---|---|
| DrishtiTable-Qwen2.5-VL-7B | Fine-tuned (SFT) | 83.2% | 89.7% |
| Claude Sonnet 4.6 | Zero-shot | 77.3% | 89.2% |
| Claude Opus 4.8 | Zero-shot | 75.5% | 88.2% |
| GPT-4o | Zero-shot | 71.1% | 84.3% |
| GPT-5.1 | Zero-shot | 69.9% | 83.3% |
| Gemini 3.1 Pro | Zero-shot | 65.7% | 73.6% |
| Fine-tuned 7B vs best zero-shot | +5.9 pp | ||
135-table held-out test set, frozen v1.0 (SHA-256), TEDS scored against private ground truth. Full 10-model board, methodology, and submission guide: live leaderboard · benchmark dataset.
Held-out multiple-choice questions. GRPO / RLVR post-training vs the open base model.
| Model | Method | Accuracy |
|---|---|---|
| Nalanda Qwen 7B | GRPO / RLVR | 66.8% |
| Qwen2.5-7B-Instruct | Base (no RLVR) | 60.5% |
| Improvement from RLVR post-training | +6.3 pp | |
Evaluated on 800 held-out MCQs from NalandaJEENEETBench. Figures are indicative and are refreshed as the benchmark split is finalized.