Close Menu
  • AI
  • Content Creation
  • Tech
  • Robotics
AI-trends.todayAI-trends.today
  • AI
  • Content Creation
  • Tech
  • Robotics
Trending
  • Potential Cognition AI Spherical Would Elevate $2B-Plus at $75B Valuation – Unite.AI
  • OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI
  • A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels
  • Collectively AI Discloses Immediate Leak After Coding Agent Uncovered Credential – Unite.AI
  • Dragos Finds Immediate Leaks and Stolen Keys in AI Provide Chains
  • Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof – Unite.AI
  • AWS Routes HyperPod Inference Throughout Areas From One Endpoint – Unite.AI
  • DailyAI Briefing: A Lean-checked Riemann proof, plus 4 extra
AI-trends.todayAI-trends.today
Home»Tech»Kimi K3: Benchmarks, license, and serving cost comparison of DeepSeek V4Pro vs. the GLM-5.

Kimi K3: Benchmarks, license, and serving cost comparison of DeepSeek V4Pro vs. the GLM-5.

Tech By Gavin Wallace19/07/20266 Mins Read
Facebook Twitter LinkedIn Email
LifelongAgentBench: A Benchmark for Evaluating Continuous Learning in LLM-Based Agents
LifelongAgentBench: A Benchmark for Evaluating Continuous Learning in LLM-Based Agents
Share
Facebook Twitter LinkedIn Email

The top three Chinese laboratories now dominate the leaderboard for open weight. Moonshot’s Kimi K3, DeepSeek V4 ProZhipu AI is a new AI technology GLM-5.2 These models all use sparse Mixture-of-Experts(MoE), with windows of a million tokens. They all target long-horizon agent and coding workloads. The article compares the three factors that an AI team will actually consider: measuring capability, licensing terms and cost.

‘Trillion-parameter’ fits Kimi K3 (2.8T) and DeepSeek V4 Pro (1.6T). GLM-5.2 has 744B in total and is therefore the smallest model of all three. The reason it is ranked first in the open-weight category before K3 arrived, is that GLM-5.2 was at the forefront of this field.


The top three candidates

Kimi K3 is a 2.8-trillion-parameter Stable LatentMoE model activating 16 of 896 experts per token. Moonshot does not know the exact number of parameters active. K3 has native vision as well as a 1M token context and reasoning always-on. Moonshot refers to it as the world’s first 3T open-class vehicle. We have covered the launch of this vehicle. here.

DeepSeek V4 Pro is a 1.6-trillion-parameter MoE with 49B active parameters, using 384 routed experts plus one shared expert. This MoE has a context window of 1M tokens with a 384K maximum output. V4 Flash is a smaller variant that covers lower workloads. The weights have been added. Hugging Face.

GLM-5.2 is a 744-billion-parameter MoE with roughly 40B active parameters and a 1M-token context window. Zhipu includes High and Max reason modes. API Access is Included.

Special Kimi K3 DeepSeek V4 Pro GLM-5.2
Total Parameters 2.8T 1.6T 744B (753B per Artificial Analysis)
Active parameters No disclosure (16/896 experts). 49B ~40B
Context window 1M Maximum output: 1M (maximum 384K). Maximum output: 1M (131K).
Modality Video + text + image Text Text
Released August 16, 2026 April 24, 2020 June 13, 2026

Benchmarks

The vendor-reported score is based on different harnesses. As a result, the benchmark numbers are not always consistent across laboratories. The neutral comparator will be the Artificial Analysis Intelligence IndexThe score is the highest when all three are on the same suit.

This index is a good place to start. Kimi K3 scores about 57DeepSeek V4 Pro scored 44 points, while GLM-5.2 received 51. K3 is ranked #3 in the overall rankings, just behind Claude Fable 5, GPT-5.6, and Opus 4.8. It’s comparable with GPT 5.5 and Opus 4.8. GLM-5.2 occupied the number one open-weight position until K3 was shipped.

Code benchmarks are similar, with a few caveats. Moonshot runs both K3 and the GLM-5.2 harnesses through its own table. On every benchmark, K3 is far ahead of GLM-5.2.

Benchmark (Moonshot harness) Kimi K3 GLM-5.2
DeepSWE 67.5 46.2
Program Bench 77.8 63.7
Terminal Bench 2.1 88.3 82.7
FrontierSWE 81.2 67.3
SWE Marathon 42.0 13.0
Automation Bench 30.8 12.9
GPQA-Diamond 93.5 91.2

DeepSeek is not included in Moonshot, therefore its results are based on separate tests. DeepSeek-V4-Pro-Max scores 80.6% on SWE-bench VerifiedThe Gemini Pro 3.1 Pro was the top-rated open weight at its launch. This model also scores 83.5 in MRCR 1M proving its ability to work well over a long period of time. GLM-5.2 was scored 62.1 on SWE-bench ProGPT-5 is a close second at 58.5.

In terms of measured ability, K3 wins out. DeepSeek V4 Pro can compete on isolated tasks of coding. GLM is a little behind the K3, but it’s still an excellent open-weight alternative.

License

The three models are all open-weight, but their practical application is different today.

DeepSeek V4 Pro is MIT-licensedHugging Face is a product that has weights from the very first day. GLM-5.2 is also MIT-licensedHugging Face, under zaiorg.org. Full weights. They both allow commercial use without restrictions, fine tuning, and self hosting.

Kimi K3 represents the exception. Moonshot committed to publish the weights before July 27, 2026. This is under a Modified MIT licence. K3 will be available only via the API or Kimi applications until that time. Moonshot’s Modified MIT Terms, which were recently updated, add a clause on attribution. The trigger is only activated if there are more than 100M monthly active users.

Cost of serving

The API List Pricing is a major difference between these two models.

Model The Input Rate ($/MTok). Input (Mtoks) Cached input
Kimi K3 3.00 15.00 0.30
DeepSeek V4 Pro 0.435 0.87 ~0.0036
GLM-5.2 1.40 4.40 0.26

DeepSeek V4 Pro has the lowest cost by far. A dollar will buy approximately 1.15M outputs tokens for V4 Pro. GLM-5.2 is about 227K and K3 about 67K.

Artificial Analysis’ pricing is based on a blended cache/input/output ratio of 7:2 to 1, which eliminates the vendor frame. According to this source, K3 is priced at $2.31 for 1M tokens. GLM 5.2 comes in at $0.90. DeepSeek V4 pro costs $0.18. The same source also reports cost per task. K3 at $0.94, GLM-5.2 at $0.32, and DeepSeek V4 Pro at $0.04.

The speed of the vehicle also varies. Artificial Analysis is a measure. GLM-5.2 at about 168 tokens/secMoonshot’s cache hit rate is above 90% in the coding workloads. This reduces K3 input costs to $0.30 for every million. Moonshot reported above 90% caching hits for coding workloads. This reduces K3’s input costs to $0.30 per $1,000,000.

The self-hosting constraint is different. GLM-5.2 744B at BF16 requires more than 1TB in VRAM, which is roughly equivalent to 8xH200 at FP8. DeepSeek V4 Pro at 1.6T needs more still. Kimi K3 has the most weight: Moonshot suggests 64 accelerators or more, which is out of reach for many teams. K3 is using MXFP4 activations with MXFP8 weights for greater hardware support.



The right model for the job

DeepSeek V4 Pro offers the lowest token cost at a high coding standard. DeepSeek V4 Pro’s license is free, it offers downloadable weights, and the output price beats its competitors.

Kimi K3 has the best measured capabilities, but costs between 5x and 17x as much to produce. Weights will not be available for download until 27th July. GLM-5.2 is in the middle: it’s cheaper than K3, quicker than either rival, can be self-hosted today and has more capability than its small size would suggest.

DeepSeek GLM is the best choice if you want to make a decision based on license transparency and verification depth. Buyers who are chasing the best benchmark scores should wait for K3 or pay API Premium.

What you need to know

  • Kimi K3 has the highest Artificial Analysis Intelligence Index score (57; #3 in total) and is only using APIs until July 27,
  • DeepSeek V4 is the leader in cost: only $0.04 for each task, and about 1.15M output tokens to every dollar.
  • GLM-5.2 (744)B is the smallest, yet fastest machine today. It can be self-hosted under MIT.
  • Only DeepSeek GLM and GLM offer open weights.


Michal is a professional in data science with a Masters of Science degree from the University of Padova. Michal Sutter excels in converting complex datasets to actionable insights. He has a strong foundation in statistics, machine learning and data engineering.

ar Benchmark deepseek
Share. Facebook Twitter LinkedIn Email
Avatar
Gavin Wallace

Related Posts

A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels

22/11/2026

OpenAI Ships GPT-Stay-1.1 for Voice Brokers, Cuts Idle Value to $0.02/Min

21/11/2026

Meet Pondering Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac

21/11/2026

Baseten Releases Bison-1 and Groundhog-1, Topping Choice Mannequin Index

21/11/2026
Top News

DailyAI Briefing: Lawsuit goals to halt new AI fashions, plus 4 extra

OpenAI Models Hacked and Escaped Containment

Anthropic Sues Department of Defense for Supply Chain-Risk Determination

When You Ought to—and Shouldn’t—Use AI to Search the Internet

Roblox’s AI-Powered Age Verification Is a Complete Mess

Load More
AI-Trends.Today

Your daily source of AI news and trends. Stay up to date with everything AI and automation!

X (Twitter) Instagram
Top Insights

AI Agents are too cheap for our own good

12/06/2025

How Great Product Teams Decide What to Build Next – Unite.AI

08/08/2026
Latest News

Potential Cognition AI Spherical Would Elevate $2B-Plus at $75B Valuation – Unite.AI

22/11/2026

OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI

22/11/2026
X (Twitter) Instagram
  • Privacy Policy
  • Contact Us
  • Terms and Conditions
© 2026 AI-Trends.Today

Type above and press Enter to search. Press Esc to cancel.