Key Takeaways
- Kimi K3 tops the Artificial Analysis Intelligence Index at 57, but is API-only with weights coming July 27.
- DeepSeek V4 Pro offers the lowest serving cost at $0.04 per task and is immediately available under MIT license.
- GLM-5.2 is the fastest at 168 tokens/sec and self-hostable today, making it a balanced open-weight option.
The Open-Weight Arms Race Reaches Trillion-Parameter Scale
Open-weight AI models have entered a new tier. This week, three Chinese labs — Moonshot AI, DeepSeek, and Zhipu AI — released trillion-parameter sparse MoE models that now top the open-weight leaderboard. Each model brings million-token context windows and targets long-horizon coding and agent workloads. But behind the impressive parameter counts lie stark differences in capability, license, and cost. We break down what matters for AI teams deploying at scale.
Table of Contents
The Contenders
Kimi K3 is Moonshot AI’s 2.8-trillion-parameter Stable LatentMoE model, activating 16 of 896 experts per token. It supports text, vision, and video with a 1M-token context window. DeepSeek V4 Pro is a 1.6-trillion-parameter MoE with 49B active parameters, using 384 routed experts plus one shared expert. It offers a 1M-token context window and 384K max output. GLM-5.2 from Zhipu AI is the smallest at 744B total parameters with roughly 40B active, but it earns its place by leading open-weight intelligence before K3 shipped.
Benchmark Analysis
Vendor-reported benchmarks are tricky due to different harnesses. The neutral comparator is the Artificial Analysis Intelligence Index, which scores all three on the same suite. On that index, Kimi K3 scores about 57, ranking #3 overall behind only Claude Fable 5 and GPT-5.6 Sol. GLM-5.2 scores 51, and DeepSeek V4 Pro scores 44.
Coding benchmarks further differentiate them. On Moonshot’s own harness, K3 leads GLM-5.2 on every shared benchmark by wide margins — DeepSWE 67.5 vs 46.2, Program Bench 77.8 vs 63.7. DeepSeek V4 Pro, tested separately, scores 80.6% on SWE-bench Verified, tied with Gemini 3.1 Pro. GLM-5.2 scores 62.1 on SWE-bench Pro, edging GPT-5.5 at 58.6.
License and Access
License terms shape deployability. DeepSeek V4 Pro is MIT-licensed with weights on Hugging Face from day one. GLM-5.2 is also MIT-licensed under the zai-org organization. Both allow unrestricted commercial use and self-hosting now. Kimi K3 is the exception: Moonshot has committed to publishing weights by July 27, 2026, under a Modified MIT license with an attribution clause triggering above 100M monthly active users. Until then, K3 is API-only.
Serving Cost and Speed
API list pricing separates these models sharply. DeepSeek V4 Pro is the cost leader at $0.435 per million input tokens and $0.87 per million output. GLM-5.2 costs $1.40 input and $4.40 output. Kimi K3 is the most expensive at $3.00 input and $15.00 output. On a blended basis from Artificial Analysis, K3 costs $2.31 per million tokens, GLM-5.2 $0.90, and DeepSeek V4 Pro $0.18.
Speed also differs. GLM-5.2 achieves about 168 tokens per second, well ahead of DeepSeek V4 Pro and Kimi K3 at about 62 each. Moonshot reports above 90% cache hits in coding workloads, which can lower effective costs for K3. Self-hosting remains a constraint: GLM-5.2 needs over 1TB of VRAM, DeepSeek V4 Pro more, and Kimi K3 recommends 64+ accelerators.
Strategic Implications
The release of these three models reshapes the open-weight hierarchy. According to data from Artificial Analysis, Kimi K3 offers the highest measured intelligence but at a significant cost premium and with delayed weight availability. DeepSeek V4 Pro provides immediate usability and the lowest cost per task, reported at $0.04 per task by Artificial Analysis. GLM-5.2 sits as a balanced middle ground, offering competitive intelligence, fast inference, and full self-hosting capabilities today.
For AI teams evaluating deployment, the choice hinges on priorities. If budget and speed are critical, DeepSeek V4 Pro or GLM-5.2 are the immediate picks. If peak benchmark performance matters and the API premium is acceptable, Kimi K3 is the leader. However, the lack of downloadable weights until July 27 locks K3 into a controlled ecosystem, making it less flexible for enterprises requiring auditability and customization.
The open-source community now has two fully open models and one promised gateway to the highest-performing open-weight system yet. As noted in recent discussions on platforms like Reddit, the transparency of DeepSeek and GLM fosters community trust and rapid iteration. Kimi K3’s eventual release under a Modified MIT license will likely drive widespread adoption, but the delay creates a window for rivals to capture mindshare.
Conclusion
The trillion-parameter open-weight era is here. Kimi K3, DeepSeek V4 Pro, and GLM-5.2 each offer compelling but distinct value propositions. As the landscape rapidly evolves, staying informed on capability, cost, and license terms is critical; visit the detailed breakdown on MarkTechPost for further analysis.
Staying ahead in the rapidly shifting landscape of AI requires precision. To future-proof your digital strategy and scale effortlessly, you need a foundation built on precision. Optimize your site with advanced speed engineering, secure your infrastructure in high-performance hosting environments, and streamline your entire workflow through autonomous AI pipelines. If you are ready to elevate your systems, Connect with Andres at Andres SEO Expert to build your ultimate architecture.
Frequently Asked Questions
What are the key differences between Kimi K3, DeepSeek V4 Pro, and GLM-5.2?
Kimi K3 is a 2.8-trillion-parameter MoE model with 1M context window, scores #3 on AI Index (57), but has API-only access until July 2026. DeepSeek V4 Pro is 1.6 trillion parameters, MIT-licensed, cheapest API cost ($0.18/M tokens), and scores 44 on AI Index. GLM-5.2 has 744B total parameters, MIT-licensed, fast inference (168 tok/s), scores 51 on AI Index, and is self-hostable today.
Which model has the best benchmark performance?
On the neutral Artificial Analysis Intelligence Index, Kimi K3 leads with a score of 57, followed by GLM-5.2 at 51 and DeepSeek V4 Pro at 44. On coding benchmarks, K3 dominates GLM-5.2 on shared tests (e.g., DeepSWE 67.5 vs 46.2), while DeepSeek V4 Pro scores 80.6% on SWE-bench Verified, tied with Gemini 3.1 Pro.
What licenses do these models use?
DeepSeek V4 Pro and GLM-5.2 are both MIT-licensed, allowing unrestricted commercial use and self-hosting immediately. Kimi K3 will be released under a Modified MIT license with an attribution clause for >100M MAU, but weights are not available until July 27, 2026; currently API-only.
How do API costs compare across these models?
DeepSeek V4 Pro is cheapest: $0.435/M input, $0.87/M output, blended $0.18/M tokens. GLM-5.2 costs $1.40 input, $4.40 output, blended $0.90/M. Kimi K3 is most expensive: $3.00 input, $15.00 output, blended $2.31/M. However, K3 can have >90% cache hits in coding workloads, reducing effective cost.
Can I self-host these models, and what hardware is needed?
GLM-5.2 and DeepSeek V4 Pro have available weights and can be self-hosted. GLM-5.2 requires over 1TB VRAM, DeepSeek V4 Pro more, and Kimi K3 recommends 64+ accelerators. Kimi K3 weights are not available until July 2026.
Which model is best for coding tasks?
For peak coding benchmark performance, Kimi K3 leads on its own harness with high scores (DeepSWE 67.5, Program Bench 77.8). DeepSeek V4 Pro also performs strongly on SWE-bench Verified at 80.6%. GLM-5.2 scores 62.1 on SWE-bench Pro, slightly above GPT-5.5. The best choice depends on cost and availability needs.
When will Kimi K3 weights be released?
Moonshot AI has committed to publishing Kimi K3 weights by July 27, 2026, under a Modified MIT license. Until then, the model is only accessible via API, which limits flexibility for enterprises requiring auditability and customization.
