An in-depth analysis of a face-off between two Chinese open-source foundation models: GLM-5.3's post-training scaling path vs. Kimi K3's path of building an entirely new, massive base model. The collision between these two approaches offers one of the most revealing angles on the competition among Chinese foundation models.
GLM-5.3 (released on 2026-08-14): reuses the GLM-5.2 base (743B), with zero changes to the base model and an aggressive post-training scaling strategy. Its three levers are dozens of times more long-horizon task environments, a richer range of environment types, and an exceptionally long post-training period.
Kimi K3 (released in mid-July 2026): 2.8 trillion parameters (the world's largest open-source model), a KDA hybrid architecture, native multimodality, and a 1M context window.
| Benchmark | GLM-5.2 | GLM-5.3 | What it measures |
|---|---|---|---|
| Terminal-Bench 3.0 | 4.6 | 28.3 | Complex tasks in real terminal environments |
| DeepSWE v1.1 | 46.2 | 66.9 | Long-horizon software engineering |
| Agents' Last Exam | 23.8 | 28.5 | Cross-tool collaboration and long-horizon tasks |
| GDPval-AA v2 | — | 1769 points | Real-world knowledge work across 44 professions |
| CyberGym | 77.2% | 84.5% | White-box code review and vulnerability discovery |
Interpretation: The sixfold improvement on Terminal-Bench shows that 5.2 was almost a "half-finished product" on real terminal tasks; 28.3 means it evolved from "not usable" to "able to handle serious work." DeepSWE 66.9 is the evaluation approach that comes closest to the real working state of an "AI programmer." CyberGym's 84.5% slightly exceeds Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). GDPval's 1769 points are interpreted as "professional task execution ability emerging from programming capabilities."
Zhipu's self-assessment remains clear-eyed: its advantage in the cybersecurity chain (code review → vulnerability validation → vulnerability exploitation → a real-world attack-and-defense loop) is concentrated in the first half; it "did not claim to be No. 1 in global cybersecurity capability."
| Dimension | GLM-5.3 | Kimi K3 |
|---|---|---|
| Parameters / approach | 743B, same base model; post-training only | 2.8T, entirely new base model |
| DeepSWE | 66.9 (v1.1) | 67.5 |
| Context / modality | 1M tokens | 1M tokens, native multimodality |
| Coding strengths | Terminal / CLI agents, security auditing | Long-horizon coding, frontend generation (first in the Arena at 1679 Elo) |
| Cybersecurity | CyberGym 84.5%, compared with Mythos 5 | 16 unknown vulnerabilities discovered during training |
| Generation speed | Builds on the 5.2 foundation (116.3 tokens/s) | 40.4 tokens/s |
| API pricing | Retains 5.2's cost advantage (approximately 2.8x price gap) | Approximately $12 per million output tokens |
| Official comparison | The strongest open-source model by coding feel, close to Fable 5 | Coding surpasses Claude Opus 4.8 / GPT 5.5 |
The battle of approaches: "more scale works wonders" (larger parameter counts, newer architectures, longer context) vs. "squeezing the base model" (deeper environments and longer training). GLM-5.3 shows that the post-training methodology is reusable; in theory, once the next-generation base model arrives, it can be "squeezed" for another round.
The important nuance: Each company highlights the track where it has the strongest position (GLM reports DeepSWE v1.1, while K3 points to its lead on SWE Marathon and ProgramBench). In overall capability, no one has yet dislodged Anthropic and OpenAI from the top spots.
The old landscape before 5.3's release (third-party tests): Artificial Analysis overall intelligence—K3 (max) 60 vs. GLM-5.2 (max) 53; K3 ranked first in frontend coding at 1679 Elo; speed—GLM 116.3 vs. K3 40.4 tokens/s; GLM was cheaper across the board (approximately a 2.8x price gap); on Composio frontier coding tasks, both tied at 7/12.
On 8/13, the official DeepSeek V4 Pro was released (DeepSWE surged from Preview's 7.3 to 62.7, surpassing Opus 4.8; CyberGym and AutomationBench surpassed Fable 5), along with DeepSeek Harness v0.1 (released under the MIT license, with "everything as a plugin").
Around the same time, SpaceXAI released Grok 4.6. All four major players are betting on "coding + agents."
Capital markets remained cool: Zhipu's stock was down more than 4% when The Paper published its report.
Backend engineering / DevOps and Claude Code-style terminal-agent workflows: GLM-5.3 is a better fit (a sixfold Terminal-Bench improvement, CyberGym at 84.5%, and a first-tier global position in the first half of the security-auditing chain).
Multimodal input: Kimi K3 has native multimodality; for long documents, both offer a 1M context window.
Multi-agent systems: K3's Swarm cluster / Goal mode works out of the box; choose DeepSeek Harness if you want control and customizability.
Cost and speed: GLM's strongest cards (nearly 3x faster and approximately a 2.8x cost gap). "Choose GLM if cost matters; choose K3 if you are deliberately buying the capability ceiling—and GLM-5.3 is making that either-or choice increasingly blurry."
Fallback recommendation: Both models are open source, so deploy them locally and test them on real bug tickets and real terminal tasks.
Summary: "GLM-5.3 bought raw intelligence headroom with a new 2.8-trillion-parameter base model, while GLM-5.3 proves th… This is a necessary excerpt; read the original source for full context.
GLM-5.3
