GLM-5 review: an MIT-licensed open MoE to pick for coding
GLM-5 (glm-5) is a strong pick when you want coding and agent work on an open model with an MIT license, at a lower cost than the closed frontier. Reviews also report that it trails the top of its class on long-context reading and strict instruction following, so pick it for the jobs it is good at.
Z.ai has since shipped GLM-5.2 and GLM-5.3, and FastMetal serves both. GLM-5 is still served.
What it is
Zhipu AI (Z.ai) released GLM-5 on February 11, 2026.
- Architecture: sparse Mixture-of-Experts, about 744B total parameters, about 44B active per token
- Context: about 200K tokens (202K reported)
- Weights: open, MIT license, on Hugging Face
The license is the headline. MIT allows commercial use, modification and fine-tuning with almost no conditions, which makes GLM-5 near-frontier coding performance you can run and change yourself.
Strengths: coding and agents
GLM-5 stands out on software engineering evaluations. Several reviews report 77.8% on SWE-bench Verified, a few points behind the best closed models (table below). One report puts it at 86% on GPQA, a science benchmark.
Agent results are good too. Artificial Analysis measured an agentic index of 68.4 and a tau2 score of 0.974 for reliability on tool-using tasks. That profile suits multi-step automation that combines function calling with a fairly long context.
Weak spots
The flip side of a coding-focused model:
- Instruction following is passable without standing out (ifbench 0.552).
- Long-context reasoning is below average for its class (lcr 0.37). Don't expect careful reading of very long documents.
- Output runs long, which some users say pushes up token spend in production.
- On general tasks outside coding, reviewers still rate the closed frontier models ahead.
How it compares
SWE-bench Verified as reported by reviews (decimals vary by source):
| Model | Type | SWE-bench Verified |
|---|---|---|
| GLM-5 | Open (MIT) | 77.8% |
| Claude Sonnet 4.6 | Closed | 79.6% |
| GPT-5.2 | Closed | 80.0% |
| Claude Opus 4.6 | Closed | 80.8% |
The gap is a few points, so "an open model close to the frontier on coding" holds up. Among open models (DeepSeek, Qwen, MiniMax), GLM models are usually near the top for agentic coding, while closed models still lead on general intelligence indexes. This review covers glm-5 only.
Cost
GLM-5 is billed per million input and output tokens, and as an open model it is generally cheaper to run than the closed frontier models above. The current rate is on the model page. Because its answers tend to run long, budget for the output side when you estimate.
Outside English
GLM-5 is multilingual and gives usable answers on everyday tasks in languages like Japanese, but we found no public benchmark for them. If tone, terminology or formality matter, test it on your own prompts before you adopt it.
Where it fits
Good fit: coding, code review and multi-step agent work; automation that leans on tool calls; jobs where you want strong coding at a lower cost.
Poor fit: precise reading or summarizing of very long documents; tasks where strict instruction following comes first; workloads where every output token counts.
FAQ
Is GLM-5 an open model?
Yes. The weights are published under the MIT license, so commercial use, modification and fine-tuning are allowed.
How does it compare with closed frontier models?
On coding it comes within a few points of the top Claude and GPT models (77.8% on SWE-bench Verified). On general intelligence and long-context reasoning, reviewers agree the closed models lead.
Does it work in languages other than English and Chinese?
It handles everyday tasks in other languages, but there is no public benchmark to point to. Test on your own prompts for anything that matters.
Try it on FastMetal
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "glm-5",
"messages": [{"role": "user", "content": "Review this function for bugs and suggest a fix: ..."}]
}'
Change model to glm-5.2 or a closed model like anthropic-claude-sonnet-4-6 and the same code compares quality, speed and cost side by side. Every model is in the catalog; rates are on the pricing page.