GLM on a Single RTX 5090: Can Any Model Survive the Homelab Bakeoff?

GLM Is the New Hotness, So Let’s Test It On the Homelab

Z.ai’s GLM family is generating buzz among developers for its open weights and coding claims. On a single RTX 5090 homelab, only the 30B-A3B GLM-4.7-Flash model has a realistic path to run as a useful local agent.

Why This Matters

The GLM hype is real, but the homelab filter separates practical models from science projects. Kimi K2 technically ran but required a 579 GB download, over six minutes to load, and generated at ‘interactive-punishment speed.’ Tool-calling failures killed Devstral before it wrote a line of code—making the API translation layer the real gate, not model intelligence.

Key Insights

• GLM family spans three distinct deployment targets: the frontier-scale 753B GLM-5.2, the lightweight 30B-A3B GLM-4.7-Flash, and the small 9B GLM-4-9B-Chat baseline.
• Tool-calling is the price of admission: models that emit fake tool calls as plain text—like Devstral in Round 7—are dead on arrival, regardless of code quality.
• Local-model discourse collapses three different claims into ‘runs’: fitting entirely in VRAM with interactive response differs drastically from mmap’ing hundreds of gigabytes from NVMe with slow offload.
• The tag-manager bakeoff task exposes real failure modes: tool-call failure, repo navigation, TypeScript debugging, build-loop behavior, goal prioritization, and shipping discipline.
• Previous round results show Qwen 3.6 built the feature but burned 77 messages on screenshots, while Devstral never made a single structured tool call.

Practical Applications

• Agentic coding (Coder Agents): Use GLM-4.7-Flash for multi-file TypeScript projects where structured tool calls and build verification are required
• LLM testing: Apply the 4-gate qualification process (load, plain chat, tool call, tiny agent task) to validate any new model before committing to a full run
• Benchmark design: Reuse the tag-manager task with Playwright screenshot requirement to catch goal-prioritization failures and build-loop behavior

References:

• https://dev.to/carryologist/glm-is-the-new-hotness-so-lets-test-it-on-the-homelab-609