GLM-4-0414 (GLM-4-32B, GLM-Z1, GLM-Z1-Rumination)

2025-04-14 · Publisher : Z.ai · Domain : Artificial Intelligence

Zhipu AI released the GLM-4-0414 family under the MIT license: GLM-4-32B, the GLM-Z1 reasoning models (32B and 9B) and GLM-Z1-Rumination, presented as comparable to GPT-4o and DeepSeek-V3/R1.

### Context In April 2025, the open market is dominated by DeepSeek-V3 and R1, which are very large (671B in total according to the specialized press), and by the Qwen2.5 models. Zhipu AI, a company spun out of Tsinghua University, had released GLM-4-9B in June 2024. This new series aims at an intermediate-size model, deployable locally, with displayed scores close to much larger models. ### What the model brings - A dense 32-billion-parameter model, GLM-4-32B-0414, pre-trained on 15T tokens of data, including a significant share of synthetic reasoning data. - A post-training combining alignment on human preferences, rejection sampling and reinforcement learning for instructions, engineering code, function calling and artifact generation (web pages, SVG). - GLM-Z1-32B-0414, a reasoning model obtained through cold start, extended reinforcement learning (math, code, logic) and pairwise preference feedback. - GLM-Z1-Rumination-32B-0414, a “rumination” model (long reflection with web search) positioned against Deep Research-style in-depth search features. - A 9B variant, GLM-Z1-9B-0414, described by the vendor as the best result among open models of comparable size. ### Architecture and specifications - Dense models (not MoE): 32B and 9B, with Base, chat and reasoning versions. - Data: 15T tokens for GLM-4-32B-Base-0414. Context window and knowledge cutoff: not disclosed in the model card; the Z1 card recommends enabling YaRN beyond 8,192 input tokens. - Languages: Chinese and English. - License: MIT (commercial use allowed, with no threshold or geographic restriction). - Weights: Hugging Face (zai-org/GLM-4-32B-0414 and variants) and ModelScope. ### Cost and access | Item | Value | |---|---| | Z.ai API, GLM-4-32B-0414-128K (current pricing) | 0.1 USD input / 0.1 USD output per million tokens | | Self-hosting | Required hardware not disclosed by the vendor in the card | ### Benchmarks | Benchmark | GLM-4-32B-0414 | GPT-4o-1120 | Qwen2.5-Max | DeepSeek-V3-0324 (open) | |---|---|---|---|---| | IFEval | 87.6 | 81.9 | 85.6 | 83.4 | | BFCL-v3 (overall) | 69.6 | 69.6 | 50.9 | 66.2 | | TAU-Bench Retail | 68.7 | 62.8 | 58.3 | 60.7 | | TAU-Bench Airline | 51.2 | 46.0 | 22.0 | 32.4 | | SimpleQA | 88.1 | 82.8 | 79.0 | 82.6 | Source: Hugging Face model card GLM-4-32B-0414, scores measured by the vendor; SimpleQA and HotpotQA on about 500 cases, with search tools, average of 3 runs. On SWE-bench Verified, GLM-4-32B-0414 obtains 33.8 (Moatless), 30.7 (Agentless) and 27.2 (OpenHands) according to the same card; no closed-model score appears there. ### Versus closed models In the table published by the vendor, GLM-4-32B-0414 outperforms GPT-4o-1120 on IFEval (87.6 versus 81.9), TAU-Bench Retail (68.7 versus 62.8) and SimpleQA (88.1 versus 82.8), and matches it on BFCL-v3 overall (69.6). These comparisons are chosen and measured by Zhipu AI, with no independent reproduction cited. The model is a dense 32B, therefore runnable on a small number of GPUs, whereas the size gap with DeepSeek-R1 is presented as the main argument. For the open versus closed choice, the benefit is the MIT license, possible self-hosting and an API cost of 0.1 USD per million tokens for both input and output on the vendor's platform. In agentic code, the 33.8% on SWE-bench Verified remains well below the scores of the next generation (GLM-4.5: 64.2%). ### Key takeaways GLM-4-0414 marks Zhipu AI's return among open models usable in the enterprise: MIT license, modest size, reasoning and in-depth search included. It is the foundation of the GLM-4.1V series and then of the move to MoE with GLM-4.5.

Tags : LLM, Open Source, Zhipu AI, Z.ai, GLM, MIT, Raisonnement