Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2: Benchmarks, Costs, and Licenses Compared

Three Chinese AI labs now occupy the top of the open-weight leaderboard. Moonshot AI's Kimi K3 (released July 16, 2026), DeepSeek V4 Pro (April 24), and Zhipu AI's GLM-5.2 (June 13) are all sparse Mixture-of-Experts models with one-million-token context windows aimed at coding and agent tasks. They differ sharply on measured capability, API price, and license terms. Here is what each model actually offers and which one fits which workload.

What happened

Detail	Value
Kimi K3 total parameters	2.8 trillion (16 of 896 experts active per token)
DeepSeek V4 Pro total parameters	1.6 trillion (49B active, 384 routed experts)
GLM-5.2 total parameters	744 billion (~40B active)
Context window (all three)	1 million tokens
Kimi K3 weights available	API only until July 27, 2026
DeepSeek V4 Pro license	MIT, weights on Hugging Face
GLM-5.2 license	MIT, weights on Hugging Face

All three models use sparse MoE architecture, meaning only a fraction of parameters activate for each token. That keeps inference cheaper than a dense model of the same total size. Each targets long-context coding, agentic workflows, and reasoning tasks.

How do they compare on benchmarks?

The cleanest comparison comes from the Artificial Analysis Intelligence Index, which tests all three on the same suite. Kimi K3 scores 57, GLM-5.2 scores 51, and DeepSeek V4 Pro (Max reasoning) scores 44. K3 ranks third overall on that index, behind only Claude Fable 5 and GPT-5.6 Sol, and comparable to Opus 4.8 and GPT-5.5. GLM-5.2 held the top open-weight position until K3 launched.

Moonshot ran K3 and GLM-5.2 through matched coding harnesses. The results across shared benchmarks:

Benchmark	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 did not appear in Moonshot’s table. From separate testing, DeepSeek V4 Pro Max scored 80.6% on SWE-bench Verified, which was the highest open-weight result at its release and tied with Gemini 3.1 Pro. It also posted 83.5 on MRCR 1M, a long-context retrieval test. GLM-5.2 scored 62.1 on SWE-bench Pro, ahead of GPT-5.5 at 58.6.

What do they cost to run?

API pricing separates these models more than benchmarks do:

Model	Input ($/MTok)	Output ($/MTok)	Cached input ($/MTok)
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

At list output rates, one dollar buys roughly 1.15 million output tokens from DeepSeek V4 Pro, about 227,000 from GLM-5.2, and about 67,000 from K3. On Artificial Analysis’s blended 7:2:1 cache/input/output pricing, K3 costs $2.31 per million tokens, GLM-5.2 costs $0.90, and DeepSeek V4 Pro costs $0.18. Cost per task on that same basis: K3 at $0.94, GLM-5.2 at $0.32, and DeepSeek V4 Pro at $0.04.

Moonshot notes above 90% cache hit rates in coding workloads, which drops K3’s effective input cost to $0.30 per million. That narrows the gap somewhat for repetitive agent loops, but the output cost gap remains large.

Speed also differs. Artificial Analysis measures GLM-5.2 at about 168 tokens per second, well ahead of both DeepSeek V4 Pro and Kimi K3 at roughly 62 tokens per second each.

Self-hosting is a different constraint

GLM-5.2 at 744B requires over 1TB of VRAM in BF16, or about 8x H200 GPUs at FP8. DeepSeek V4 Pro at 1.6T needs more still. Kimi K3 is the heaviest: Moonshot recommends 64 or more accelerators, which puts local serving out of reach for most teams. K3 uses MXFP4 weights with MXFP8 activations for broader hardware support.

License terms: what you can actually do today

DeepSeek V4 Pro and GLM-5.2 are both MIT-licensed with weights available on Hugging Face now. Both allow unrestricted commercial use, fine-tuning, and self-hosting without restrictions.

Kimi K3 is the exception. Moonshot has committed to publishing weights by July 27, 2026, under a Modified MIT license. Until then it is accessible only through the API and Kimi apps. The Modified MIT terms add one attribution clause, but it only triggers above 100 million monthly active users, so it is irrelevant for almost every business.

Which model fits which job?

• Lowest cost per token with solid coding quality: DeepSeek V4 Pro. MIT weights are downloadable today, output price undercuts both rivals, and its SWE-bench score is competitive.
• Highest measured capability: Kimi K3, but at 5x to 17x the output price of DeepSeek V4 Pro, with no downloadable weights until July 27.
• Middle ground: GLM-5.2 is cheaper than K3, faster than both rivals, and self-hostable on 8x H200 hardware today.

Our take

The gap between K3’s benchmark scores and its price is real. For most business applications, including AI integration into products or workflows, the 23x output cost premium over DeepSeek V4 Pro is hard to justify unless the task specifically rewards K3’s capability ceiling. K3’s missing weights are also a concern for teams that need reproducibility or want to avoid API lock-in right now.

DeepSeek V4 Pro looks like the practical default for cost-sensitive agent and coding workloads. Its task cost of $0.04 versus K3’s $0.94 is not a rounding error. GLM-5.2 makes sense if generation speed matters (168 vs. 62 tokens/sec is a meaningful latency difference for interactive tools) and if you have the VRAM budget to self-host.

We covered the initial market reaction to Kimi K3 in our earlier piece on what Kimi K3 means for the open-source AI field. The short version: Chinese labs releasing competitive open weights at these price points continues to pressure the proprietary model market in ways that benefit buyers.

Watch the July 27 weight release. If Moonshot delivers on the Modified MIT terms as described, K3 becomes a much more credible option for teams that can provision the hardware.

Source: Marktechpost