🔍 THUDM launches GLM-4-32B: the open-weight challenger to GPT-4o and DeepSeek V3

On 14 April 2025, Chinese research lab THUDM released GLM-4-32B, a 32-billion-parameter open-weight language model.

THUDM Releases GLM 4: A 32B Parameter Model Competing Head-to-Head with GPT-4o and DeepSeek-V3

THUDM Releases GLM 4: A 32B Parameter Model Competing Head-to-Head with GPT-4o and DeepSeek-V3

MarkTechPostAsif Razzaq

Its Hugging Face profile reveals that it was pre-trained on 15T tokens of high-quality data and fine-tuned using RLHF techniques, rejection sampling, and human preference alignment.
In benchmarks like IFEval and BFCL-v3, GLM-4-32B achieves 87.6% and 69.6%, outperforming or matching GPT-4o-1120 and DeepSeek-V3-0324.
THUDM also introduced the Z1 and Z1-Rumination variants, focused on deep reasoning and investigative writing, further refining its mathematical and problem-solving capabilities.

THUDM/GLM-4-32B-0414 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

The official repository confirms an Apache-2.0 license and support for local deployment, making it easy to integrate into both private servers and edge devices via GitHub.

Why is this relevant?

1. Efficiency vs. size
   With just one-tenth the parameters of GPT-4o, it achieves similar results thanks to training and architectural optimizations.
2. Competitive timing
   OpenAI updated GPT-4o on April 25 with improvements in memory and STEM capabilities (OpenAI Help Center), while DeepSeek launched V3 in March with 37B active and 671B total parameters, focused on code and reasoning (PYMNTS.com, Redblink).
   GLM-4-32B now offers a third path: fully open and free from API costs.
3. Open ecosystem
   Since it’s available on Hugging Face and GitHub, the community can already test quantizations, fine-tune it for specific domains, and deploy it in on-prem solutions.

How to try it out

• Direct download available on Hugging Face (THUDM/GLM-4-32B-0414) with checkpoints in safetensors format.
• Inference examples using Transformers or vLLM are available in the README.
• For agent tasks, try the Z1-Rumination version, which already includes function calling and search capabilities.

Final Thoughts

The launch of GLM-4-32B confirms that 2025 is the year of the “lightweight powerhouse”: beyond parameter counts, the real game-changers are data quality, alignment, and tool-use.
With its open license and top-tier performance, GLM-4-32B is set to accelerate the adoption of private assistants, corporate agents, and edge projects where GPT-4o or DeepSeek are unfeasible due to cost or hardware requirements.

¿Will you give it a try? Let me know in the comments which use case you're interested in, and let’s share results!

References

1. MarkTechPost – launch announcement (14 Apr 2025) MarkTechPost
2. GitHub – THUDM/GLM-4 repo & license GitHub
3. Hugging Face – GLM-4-32B-0414 model card Hugging Face
4. Hugging Face – Evaluation Results (IFEval, BFCL-v3) Hugging Face
5. OpenAI Help Center – mejoras GPT-4o (25 Apr 2025) OpenAI Help Center
6. PYMNTS – DeepSeek V3 launch announcement (25 Mar 2025) PYMNTS.com
7. RedBlink – DeepSeek V3 specs (Apr 2025) Redblinkace