HKU's PrismaCoder: Open Multimodal Code Models Rival Proprietary Systems
hkunlp · hf · 2026-10-22
HKU's hkunlp team releases PrismaCoder, open multimodal code intelligence models extending code AI beyond source text to what programs render: charts, web pages, diagrams, documents, and animations. Key points: - PrismaForge: a pipeline synthesizing per-instruction rubrics with separate code-level and visual-level checks, serving as both data filter and RL reward; - PrismaCoder-2M: the broadest open multimodal code dataset — 2M samples across 19 task families (WebUI generation 17.8%, Chart-to-Code 9.8%, Image-to-SVG 9.0%, LaTeX, Manim, chemical structures), 59.4% vision-centric; - PrismaCoder-9B/35B: lead open models by a wide margin, matching or surpassing proprietary systems on charts, front-end pages, and semi-structured documents; - Analysis: rubric filtering beats direct judging, the recipe transfers across backbones, and rubric scores align with human judgment.
More from coding & agent
- I built an AI software factory that turns GitHub issues into PRs on autopilot — Pavan_Belagatti · 2026-10-22
- Dev builds a full in-game map editor for his AI-made dungeon crawler with Codex — Dimillian · 2026-10-22
- Walmart Opens MCP Agent Gateway While Amazon Locks Out AI Shopping Agents — sanjaykalra · 2026-10-22
- open-duet: open-source self-hosted Claude Duet alternative hits 5.9k GitHub stars — matchaman11 · 2026-10-22
- Locals: Open-Source Apple Silicon Engine Runs Local LLMs Up to 3x Faster — SumitGup · 2026-10-22
- When AI automation saves less than it costs: the hidden maintenance tax of agents — djgoel123 · 2026-10-22