llm - Blog

GLM-OCR is a 0.9B-parameter multimodal OCR model built on the GLM-V architecture, designed for complex document understanding, not just text extraction. Delivers structure-first outputs (semantic Markdown, JSON, LaTeX), accurately reconstructing tables, formulas, layout, and handwriting across 100+ languages with state-of-the-art OmniDocBench V1.5 performance (94.62) at ~1.86 PDF pages/second.

Qwen3-Coder-Next is an open-weight language model achieving Sonnet 4.5-level coding performance with only 3B activated parameters (80B total with MoE architecture). Runs on consumer hardware like 64GB MacBook, RTX 5090, or AMD Radeon 7900 XTX with 256K context length. Scores 44.3% on SWE-Bench Pro while eliminating expensive API costs.