Request: Replace GLM-5 with GLM-5.1 from Z-ai on NIM

AI & Data Science NVIDIA NIM Models
,
1

Hi NVIDIA team,

I’d like to request that GLM-5 currently available on the NIM model catalog be replaced or supplemented with GLM-5.1 from Z-ai, as GLM-5.1 is a significant post-training update.

Why GLM-5.1 is a Better Fit for NIM

GLM-5.1 is built on the exact same architecture as GLM-5 (744B total / 40B active MoE parameters), but has been heavily fine-tuned for agentic coding workflows. The key improvements include:

What Changed in 5.1

  1. Better Agentic Task Performance

    • Improved instruction following — stays locked onto the main objective without deviating
    • Self-debugging loops — runs linters, catches errors, and iterates until the task is genuinely complete
    • Better planning — absorbs full context before making changes instead of rushing to generate code
    • More reliable interleaved thinking for long-horizon tasks

  2. Adaptive Reasoning Efficiency

    • GLM-5 was a heavy over-reasoner, applying deep chain-of-thought even on trivial tasks
    • GLM-5.1 adaptively scales reasoning — simple tasks get fast responses, complex tasks get full thinking
    • The net effect is noticeably snappier in day-to-day use

Spec Comparison

GLM-5.1 keeps all the same core specs as GLM-5:

  • 744B total / 40B active parameters (MoE)
  • 204,800 token context window
  • 131,072 max output tokens
  • Reasoning mode, function calling, structured output, streaming

Agentic Leaderboard Ranking

  • GLM-5: No specific agentic ranking
  • GLM-5.1: 2nd overall on agentic tasks leaderboard

Use Case

For developers running agentic coding workflows inside tools like Claude Code, Cline, or Open Claw, GLM-5.1 is a clear upgrade. GLM-5 had a known tendency to overdo things and get stuck on multi-step tasks — GLM-5.1 fixes this.

Request

Could the NVIDIA NIM team please consider updating the catalog to offer GLM-5.1 (or offer both versions as options)? The model ID on Z.ai’s platform is zai/glm-5.1.

Trade-offs to Consider

It’s worth noting that GLM-5.1 shows slight regressions in general chat/math QA and may produce code-biased outputs for plain-text questions. However, for the primary NIM use case of API-driven agentic tasks, GLM-5.1 is demonstrably superior.

Thanks for considering this request!

References:

Tags: glm, model request, upgrade, z-ai, agentic

2

GLM-5.1 isn’t even publicly available outside of their subscription based platform yet…

1 Like
3

Getting the request in before it’s even released.