Meituan's LongCat-2.5 tops 3DCodeBench at 55.4, open-sourced under MIT
dr_cintas · x · 2026-11-11
LongCat-2.5, a 1.6T-parameter model that represents images as tokens processed in the same stream as text, excels at visual coding: show it a design and it builds the page, or it writes WebGL code for 3D scenes.
Official benchmarks:
- 3DCodeBench Text→3D: 55.4, the highest in the published table (above Claude Opus 5 at 54.3)
- Vision2Web: 78.6, near Claude Opus 5's 80.2
API pricing: $0.30/M uncached input tokens, $0.006/M cached input, $1.20/M output, plus a $1.90 50M-token launch trial pack with free cache hits. Weights are open-sourced under MIT on Hugging Face.
Related event: Meituan Open-Sources 1.6T-Parameter LongCat-2.5, Topping 3DCodeBench(5 posts)→
More from coding & agent
- Two days deep-diving agentic memory: mem0's complete guide is the best starting point — sharpeye_wnl · 2026-11-11
- Meta's open-source design system Astryx hits 20K GitHub stars, #2 trending — Vjeux · 2026-11-11
- Fable 5.2 nails four pro design tasks: 689 robot parts, 4,871 plumbing plans — FinanceYF5 · 2026-11-11
- Fable 5.2 nails four pro design tasks: 689 robot parts, 4,871 plumbing plans — FinanceYF5 · 2026-11-11
- lattedit: a free local-first browser LaTeX editor with a JSON API for agents — cneuralnetwork · 2026-11-11
- Why enterprises should add a model routing layer instead of one LLM for all use cases — kashifmanzoor · 2026-11-11