MIT student fine-tunes open Molmo 2 into Molmo2Fish for sonar fish tracking
allen_ai · x · 2026-10-14
Ai2 highlights a research case for open Molmo 2: MIT BeeryLab visiting student Kai Van Brunt used its open code and training data to adapt the general video object-tracking model for sonar footage.
- He fine-tuned Molmo 2 on sonar video, then trained on examples of tracking errors, instructions, and corrections.
- The resulting Molmo2Fish could fix some mistakes after guidance, e.g. two salmon wrongly assigned the same track.
- A showcase of how open code + open data enable bespoke academic fine-tuning.
Related event: MIT Researchers Build Molmo2Fish to Track Salmon in Sonar Video(4 posts)→
More from Models
- Gemini 4 Argon goes GA, reportedly controls full-body humanoid with dexterous hands — evijit · 2026-10-14
- TV noise crashes voice agents from 38.7% to 8.6%; full duplex is the fix — SIGKITTEN · 2026-10-14
- Gemini 4 Argon Hits 69% on HieroglyphBench as Author Jokes About Benchmark Maxxing — ShikharMurty · 2026-10-14
- Gemini 4 Argon evades CoT monitors in sabotage tests, code-diff monitoring catches it — scaling01 · 2026-10-14
- TypeSafe AI's Jev Claims 100x Faster Decisions, Can Replace 9 LLM Call Types — Arindam_1729 · 2026-10-14
- Pedro Domingos reviews OpenAI math papers: impressive if it holds up, but incoherent — pmddomingos · 2026-10-14