GLM-5: A New Era of Open-Source Agentic Engineering

Forum topic · ✨步子哥 · 2026-02-21

Summary

This forum post presents a technical deep-dive into GLM-5, the latest large language model from Zhipu AI, framed as a shift from "Vibe Coding" to "Agentic Engineering." The model totals 744B parameters with 40B/44B activated per token, built on a Mixture-of-Experts architecture with 256 experts (8 active, ~5.9% sparsity), a 202K-token context window (128K max output), and pretrained on 28.5T tokens. It adopts DeepSeek Sparse Attention (DSA) to reduce attention computation by roughly 97% (e.g., compressing ~8.2B attention pairs at 128K context to ~260M), supported by a Lightning Indexer and Multi-Token Prediction. Training reportedly ran fully on Huawei Ascend hardware with MindSpore, and weights are released under the MIT license. The post cites benchmarks including 77.8% on SWE-bench Verified, a #4 global ranking and #1 open-source ranking on Artificial Analysis, and claims GLM-5 is the first open model to exceed 50 on a composite intelligence index. It also covers the Slime asynchronous reinforcement learning framework for agentic training, token-efficiency and deployment optimizations, architectural lineage, and limitations and future directions.

This post is a comprehensive technical analysis of GLM-5, the new flagship model from Zhipu AI, positioned as the start of a new era for open-source agentic engineering.

Key points

From "Vibe Coding" to "Agentic Engineering"

The release marks a strategic shift from lightweight, intuition-driven code generation toward autonomous, multi-step engineering: self-directed planning, long-term memory, continuous learning, and end-to-end software lifecycle coverage.
Three technical enablers are highlighted:
Pretraining scale-up: total parameters grew from 355B to 744B; pretraining data increased from 23T to 28.5T tokens.
Adoption of DeepSeek Sparse Attention (DSA) for cost-efficient long-context modeling.
The Slime asynchronous reinforcement learning framework with new agentic RL algorithms, enabling learning from long-horizon interactions.

Market and open-source positioning

Ranks #4 globally and #1 among open-source models on Artificial Analysis; claimed to be the first open model to surpass 50 on a composite intelligence index.
77.8% on SWE-bench Verified.
Released under the MIT license, allowing unrestricted commercial use, modification, and redistribution.
Trained entirely on Huawei Ascend chips with the MindSpore framework, demonstrating full-stack domestic compute independence.

Specifications

| Spec | Value | |---|---| | Total parameters | 744B | | Activated parameters | 40B / 44B | | Experts | 256 (8 active, ~5.9% sparsity) | | Pretraining data | 28.5T tokens | | Context window | 202K (128K max output) | | Layers | 78 (first 3 dense FFN, remaining 75 MoE) |

Architecture innovations

MoE design: 256 expert networks with top-8 routing per token; dense FFN in the first three layers for foundational stability.
DeepSeek Sparse Attention (DSA): reduces self-attention complexity from O(L²) toward near-linear. At 128K context, full attention would require ~8.2B attention pairs; selective sparsification compresses this to ~260M pairs — a 97% compute reduction. A Lightning Indexer performs sparse selection (Top-2048 tokens).
Multi-Token Prediction (MTP) for improved decoding efficiency.

Training and infrastructure

Slime asynchronous RL framework supports agent-oriented post-training.
Full training on the Huawei Ascend + MindSpore stack.

Deployment and efficiency

The post further discusses token-efficiency optimizations, hardware adaptation, architectural lineage from earlier GLM generations, and research limitations and future directions (sections truncated in the source).

Competitive comparison (as presented in the post)

| Dimension | GLM-5 | Competitors | |---|---|---| | Total params | 744B | DeepSeek-V3.2: 685B; MiniMax-M2.5: 230B | | Active params | 40B/44B | DeepSeek-V3.2: 37B | | Context window | 202K | DeepSeek-V3.2: 128K; Claude-3.5: 200K | | License | MIT | Llama: tiered; Qwen: limited commercial | | Training compute | Huawei Ascend | Mainstream: NVIDIA GPUs |

> Note: Figures above are as reported in the original forum post; the source includes citation markers to unspecified references.