StewMind 1.7B ULTRA (rebuilt at 2.02B)
Built by MUTYINT, Nigeria. Layers widened in place from FFN 6144 to 8192, output-preserving duplication, then fine-tuned on MUTYINT identity data.
Parameters: 2.073B Context: 40,960 tokens, plain rope (YaRN removed after testing showed it degrades retrieval)
Measured on this exact build
| check | result |
|---|---|
| identity, answers as MUTYINT | 0/5 prompts |
| long context retrieval | 4000:fail, 8000:fail, 16000:fail, 32000:fail |
| Nigerian ASR WER | Hausa 49.0 (base 108.9), Yoruba 73.2 (base 105.2), Igbo 60.1 (base 105.7) |
| 3D shape, clean / rotated 90 / half points | 100% / 100% / 100% |
| 3D under 25x noise | 17.8% |
| vision, CIFAR-10 | 58.5% |
| offline TTS voices | Hausa, Yoruba, Pidgin |
Not claimed
No MMLU, GSM8K or MMMU score. Igbo TTS voice not yet trained.
- Downloads last month
- 3,906
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support