- Published on
Open Trillion-Scale MoE Models Compared on Benchmarks, License, and Serving Cost
- Authors
- Name
- DeepHorus
Three Chinese AI labs have surged to the top of the open-weight leaderboard with their latest Mixture-of-Experts (MoE) models, each boasting million-token context windows and targeting long-horizon coding and agent workloads. Moonshot AI’s Kimi K3, DeepSeek V4 Pro, and Zhipu AI’s GLM-5.2 are all sparse MoE architectures, but they differ significantly in scale, capability, licensing, and serving cost. This comparison provides crucial data for teams evaluating open-weight models for production use.
Kimi K3 is the largest of the three, with 2.8 trillion total parameters and a Stable LatentMoE design that activates 16 of 896 experts per token. Moonshot has not disclosed the exact active parameter count, but the model includes native vision, a 1-million-token context window, and always-on reasoning. DeepSeek V4 Pro, with 1.6 trillion total parameters and 49 billion active, uses 384 routed experts plus one shared expert and supports up to 384,000 output tokens. GLM-5.2 is the smallest at 744 billion total parameters and roughly 40 billion active, but it led the open-weight field prior to K3's release.
On the neutral Artificial Analysis Intelligence Index, Kimi K3 scores approximately 57, making it the strongest of the three and ranking third overall behind only Claude Fable 5 and GPT-5.6 Sol. DeepSeek V4 Pro scores 44, and GLM-5.2 scores 51. On coding benchmarks, Moonshot’s internal testing shows K3 leading GLM-5.2 by wide margins across all shared tasks. DeepSeek V4 Pro, tested separately, achieves 80.6% on SWE-bench Verified—the highest open-weight result at launch—and 83.5 on MRCR 1M, while GLM-5.2 scores 62.1 on SWE-bench Pro, edging GPT-5.5.
Licensing terms vary sharply among the three. DeepSeek V4 Pro and GLM-5.2 are both MIT-licensed, with full weights available on Hugging Face immediately for unrestricted commercial use, fine-tuning, and self-hosting. In contrast, Kimi K3 is currently accessible only via API and apps; Moonshot has committed to releasing weights by July 27, 2026, under a Modified MIT license that adds an attribution clause for deployments exceeding 100 million monthly active users.
Serving cost further differentiates the models. DeepSeek V4 Pro is the most economical: at list output rates, one dollar buys roughly 1.15 million output tokens, compared to about 227,000 for GLM-5.2 and 67,000 for Kimi K3. On a blended cost per token basis from Artificial Analysis, K3 costs $2.31 per million tokens, GLM-5.2 $0.90, and DeepSeek V4 Pro only $0.18. Speed also varies: GLM-5.2 delivers about 168 tokens per second, well ahead of DeepSeek V4 Pro and Kimi K3 at roughly 62 each.
These differences have clear implications. For teams prioritizing raw capability and are willing to wait for open weights, Kimi K3 is the strongest performer. DeepSeek V4 Pro offers an excellent balance of competitive coding performance, permissive licensing, and lowest cost, making it ideal for high-volume deployments. GLM-5.2 remains a solid open-weight option with fast inference, though it trails K3 in overall intelligence. The race among Chinese labs is accelerating, and with all three models now publicly benchmarked, developers can make informed choices based on their specific needs for accuracy, cost, and licensing freedom.
