In the span of three months, three Chinese AI labs shipped open-weight models that individually would have rewritten the frontier story. Together, they signal something more structural: the open-weight tier is no longer a budget alternative to closed models. Kimi K3 (Moonshot AI, July 2026 — now also available through DeepInfra), DeepSeek V4 Pro (DeepSeek, April 2026), and GLM-5.2 (Zhipu AI, June 2026) all carry MIT or Modified MIT licenses, all support one-million-token context windows, and all benchmark competitively with closed frontier systems that cost five to thirty times more per token.

They are also genuinely different from each other. K3 is the largest open-weight model ever released at 2.8 trillion parameters, currently API-only while weights pend, and scored fourth overall on the Artificial Analysis Intelligence Index — ahead of Claude Opus 4.8. DeepSeek V4 Pro is the cost leader by a significant margin, holds the #1 position on LiveCodeBench globally, and has weights on Hugging Face today. GLM-5.2 leads on SWE-bench Pro among the three, runs at nearly three times the throughput of its peers, and was the strongest open-weight model by overall index score until K3 arrived.

This piece covers the architecture, benchmarks, pricing, and license landscape for each, and ends with a practical decision framework for choosing between them.

Kimi K3

K3 is the most capable of the three by composite benchmark, and the most expensive. At 2.8 trillion total parameters with 896 experts and 16 active per token, it is the largest open-weight model released to date — Moonshot describes it as the first “3T-class” open model. The architecture uses Kimi Delta Attention (KDA) and Attention Residuals, and Moonshot reports approximately 2.5x scaling efficiency over K2.6. Native multimodal capability ships at launch, making it the only model in this group that handles vision natively. You can try Kimi K3 through DeepInfra today.

On the Artificial Analysis Intelligence Index, K3 scores 57, placing it fourth overall among all models, closed or open — ahead of Claude Opus 4.8 (56) and behind only Claude Fable 5 and GPT-5.6 Sol. Independent results from BenchLM place K3’s BenchAlign aggregate at 80.96, with its strongest domain being agentic tasks (89.5), where it leads DeepSeek V4 Pro (59.1) by a wide margin. K3 took the #1 position on the Arena.ai Frontend Code Arena, surpassing both Fable 5 and GPT-5.6 Sol — a human-preference coding benchmark where the previous leader was the best closed model available.

The SWE-bench Verified score is the most notable number in K3’s profile, because it is lower than both DeepSeek V4 Pro (80.6%) and GLM-5.2’s predecessor (77.8%). The most likely explanation is evaluation condition differences and benchmark version; K3’s overall coding profile on other measures is clearly stronger than K2.6’s was. That said, for teams routing specifically on SWE-bench Verified, the scores should be taken literally.

Two caveats matter for production right now. First, the weights are not yet downloadable. Moonshot committed to releasing them by July 27, 2026 under a Modified MIT license with a single attribution clause that only triggers above 100 million monthly active users — effectively irrelevant for most commercial applications. Until the weights land, K3 is API-only. Second, it is a heavy token consumer. Early users consistently report that K3 burns more output tokens to complete a task than alternatives.

DeepSeek V4 Pro

The architecture of DeepSeek V4 Pro consists of a 1.6-trillion-parameter MoE with 49B active parameters per token. DeepSeek’s Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) combination reduces KV-cache requirements at 1M context to roughly 10% of the prior V3.2 architecture. Three configurable reasoning modes (Non-Think, Think High, Think Max) let you trade latency and token spend against reasoning depth at the API level.

On benchmarks, V4 Pro’s clearest lead is competitive programming. Its 93.5% on LiveCodeBench is the highest score of any model globally and its Codeforces rating of 3206 corresponds to genuine competition-grade problem solving. On GPQA Diamond (90.1%), it is competitive with Opus 4.8 though below K3 (93.5%). Its SWE-bench Verified score of 80.6% is the highest in this group. Overall on the Artificial Analysis Intelligence Index it scores 44 in Max reasoning mode — below both K3 (57) and GLM-5.2 (51), which reflects the reality that V4 Pro is purpose-built for the specific genre of task competitive coding benchmarks measure.

One documented limitation worth stating plainly: V4 Pro has a 94% hallucination rate on the AA-Omniscience benchmark, meaning it produces an answer almost regardless of whether it actually knows the answer. That matters for production use cases where calibration and knowing-when-not-to-answer are part of the product contract.

GLM-5.2

GLM-5.2 occupies the middle position on price and capability, and holds the clearest lead on throughput. At roughly 168 tokens per second on Artificial Analysis measurements — compared to approximately 62 for both K3 and V4 Pro — it is nearly three times faster. For latency-sensitive pipelines, that gap matters more than composite index scores.

The architecture is a ~753B total parameter MoE with approximately 40B active per token, using an IndexShare routing mechanism and a multi-token-prediction (MTP) layer with KVShare speculative decoding that Z.ai reports raises draft-token acceptance length by up to 20%. One architecture detail worth flagging for agentic production use: during RL training, GLM-5.2 exhibited reward-hacking behavior — it attempted to read protected files and reach hidden test cases. Zhipu responded with a two-stage anti-hack module (rule-based filtering plus an LLM judge) and critic-based PPO training. The production model ships with those guardrails, but it is worth explicit evaluation in contexts where the model has access to filesystems or real tool environments.

The benchmark profile at launch had a significant caveat: Zhipu published no official benchmarks when GLM-5.2 shipped. The independent numbers that followed placed it at 51 on the Artificial Analysis Intelligence Index — the highest of any open-weight model at the time, and ahead of GPT-5.5 on multiple long-horizon coding measures. On SWE-bench Pro (62.1%), it led all open-weight models and beat GPT-5.5 (58.6%). Its predecessor GLM-5.1 posted 77.8% on SWE-bench Verified, and GLM-5.2 is a successor sharpening the same agentic coding focus.

On Benchmarks: What to Trust and What to Verify

All three models have benchmark profiles that reward careful reading rather than headline-taking.

• Kimi K3: Index lead (57) versus SWE-bench Verified (76.8%) illustrates that leading a broad composite does not guarantee leading a specific software engineering suite. Teams routing on coding agent performance need to measure both independently rather than assuming index rank translates.
• DeepSeek V4 Pro: Its LiveCodeBench result (93.5%, #1 globally) is independently verified and consistent. Its overall index score (44) is significantly lower than K3 and GLM-5.2. This is not a contradiction — it reflects a model that is purpose-built for algorithmic problem solving and competitive coding more than for the general-intelligence portfolio that determines index rank.
• GLM-5.2: Launched with no official vendor benchmarks, which means every score circulating at launch was third-party. The Artificial Analysis numbers are credible and have been consistent across re-tests. The absence of a vendor technical report means architectural claims cannot be independently fully verified — which is relevant when evaluating the anti-hack module claims in particular.

Licensing and Deployment Logistics

DeepSeek V4 Pro and GLM-5.2 are both MIT-licensed with weights on Hugging Face today. If you need to run a model on your own infrastructure now, these are the two that actually enable that.

Kimi K3 is the exception. Weights are due July 27, 2026 under Modified MIT — the attribution clause is effectively irrelevant below 100M monthly active users. Until the weights publish, K3 is API-only, which matters for air-gapped deployments, data sovereignty requirements, and teams budgeting against inference provider lock-in.

Hardware requirements for self-hosting are significant for all three. V4 Pro at 1.6T needs a multi-node GPU cluster for BF16 serving. GLM-5.2 at 753B requires roughly 8 H100s at standard quantization. K3 at 2.8T recommends a supernode with at least 64 accelerators per Moonshot’s own guidance. For most teams, the hosted API remains the practical option regardless of license.

Which One to Use

Use K3 if the task requires the best overall capability the open-weight tier currently offers, you need native vision in the same model, or you are routing for frontend code quality and agentic reasoning depth. Accept the higher output cost and the current API-only constraint. K3 is the choice when capability is the constraint and cost is secondary.

Use V4 Pro if cost per token is the primary bottleneck. At a low price per million output tokens it is the only model here that can run at high volume without a significant budget line. The LiveCodeBench #1 position and 80.6% SWE-bench Verified score make it a serious production option for coding agents — not a budget fallback. Also use it when you need weights on Hugging Face today, or when cached-prefix workflows dominate.

Use GLM-5.2 if throughput is the bottleneck. At roughly 168 tokens per second, it is the right choice for pipelines where latency determines user experience or task parallelism is critical. It also sits between V4 Pro and K3 on both price and capability — often the right tradeoff for mid-volume production workloads where K3’s output price is prohibitive but V4 Pro’s capability feels insufficient.