Daily Model Scout Report — 2026-04-08
#7
by msudharsanan - opened
Daily Model Scout Report — 2026-04-08
Method: Queried HF image-text-to-text sorted by createdAt for the last ~24h. Most new uploads are Gemma-4 RP/abliterated derivatives and Qwen3.5 quant repacks — filtered out. Current bar: qwen3-vl-8b-sft+grpo @ 0.9131 weighted (3,500 hard eval).
Findings
| Model | Size | Family | Notes | Relevance |
|---|---|---|---|---|
| Vastined/Step3-VL-10B-GGUF | 10B | StepFun Step3-VL | GGUF repack — competitive multimodal benchmarks, fits 98GB easily. Worth a base eval. | Medium |
| Shockt/moondream-2b-2025-04-14-4bit | 2B 4-bit | Moondream | New snapshot. Small/fast lane vs qwen3-vl-2b-sft-grpo-v9 (0.8948). | Medium |
| JusMe/FoodExtract-Vision-SmolVLM2-500M-fine-tune | 500M | SmolVLM2 | Domain attribute-extraction fine-tune — recipe reference for 9-field JSON. | Low |
| seulaugues/nanoVLM | tiny | nanoVLM | Research/toy. | Low |
| thkim0305-SNU/llama3.2_3B_vl_* | 3B | Llama3.2-VL research | Academic ablations. | Low |
No High-relevance releases today. No new Qwen-VL, InternVL, Florence, PaliGemma, MiniCPM-V, or fashion-specific base models in this window. Step3-VL-10B is the only thing worth a base-model sanity eval against the hard set.