Bilingual chat models with tool use, no GPU farm required

What it does

ChatGLM3-6B is a family of open-weights bilingual (Chinese and English) dialogue models from Zhipu AI and Tsinghua’s KEG Lab. This repository hosts inference code, prompt utilities, and composite demos for four variants—standard 8 k context, a raw base model, and long-context 32 k and 128 k versions—intended for local or Hugging Face–based deployment.

The interesting bit

The base model claims the top spot among sub-10 B parameter models on the authors’ benchmark suite (GSM8K, MATH, BBH, and others), yet the chat variant still squeezes in native function calling, code interpretation, and agent-style tasks through a redesigned prompt format rather than relying on external orchestration.

Key highlights

• Four model flavors: standard ChatGLM3-6B (8 k), ChatGLM3-6B-Base, and long-context ChatGLM3-6B-32K / 128K for document-heavy workloads.
• Reported benchmarks show the base model scoring 72.3 on GSM8K and 66.1 on BBH, up sharply from the prior generation’s 32.4 and 33.7.
• A composite demo bundles chat, tool-use, and code-interpreter modes; the model executes code inside a Jupyter environment to answer complex questions.
• Weights are fully open for academic use and available for commercial use after a registration questionnaire.
• Broad ecosystem support: community ports cover chatglm.cpp quantization, TPU edge inference, TensorRT-LLM, OpenVINO, plus RAG and fine-tuning frameworks.

Caveats

• The README prominently steers users toward the newer GLM-4 and GLM-4-9B series, suggesting ChatGLM3 is no longer the flagship.
• Benchmark comparisons use a “Best Baseline” cutoff from October 2023, so the sub-10B landscape has likely shifted since then.
• The authors explicitly disclaim responsibility for model outputs and warn that the 6 B size limits output reliability.

Verdict

A solid candidate if you need a lightweight, Chinese-English conversational model with built-in tool support for local hardware. Skip it if you are already targeting the newer GLM-4 family or need independently verified, current leaderboard standings.