Kimi K3 vs GLM-5.2 vs DeepSeek V4 Pro: Which to Run

Kimi K3 is a 2.8-trillion-parameter Mixture-of-Experts model that fires just 16 of its 896 experts per token. Moonshot has not disclosed the exact active-parameter count, which is worth flagging up front because active params, not total params, drive serving cost and hardware footprint. A 2.8T total with a sparse 16-of-896 routing pattern is a much smaller compute bill per token than the total suggests, but without the disclosed active count you are estimating rather than measuring.

The architecture is where K3 gets genuinely new. It uses Kimi Delta Attention (KDA), a hybrid linear attention mechanism, paired with a 1-million-token context window. Moonshot reports that KDA delivers up to 6.3x faster decoding in million-token contexts, that a technique it calls Attention Residuals (AttnRes) gives roughly 25% higher training efficiency, and that K3 achieves roughly 2.5x better overall scaling efficiency than Kimi K2. Those are vendor-reported figures on a vendor-served model, so treat them as the ceiling, not the floor, until the weights are public and someone outside Moonshot can reproduce them.

The license is a modified MIT license carrying an attribution clause that only bites services with 100 million or more monthly active users. For essentially every team reading this, that is functionally MIT: free commercial use, free fine-tuning, free self-hosting, with an attribution string that only matters at hyperscaler traffic. The catch is not the license. It is that the thing the license governs, the weights, does not exist in public yet.

Let me separate what K3 genuinely leads on from what the numbers hide, because the Kimi K3 benchmarks coding story is stronger than the caveats and the caveats matter more than the marketing admits.

On the Arena Frontend Code leaderboard, K3 took the #1 spot with 1,679 points, ahead of Claude Fable 5 (1,631), GPT-5.5 (1,618), and GLM-5.2 (1,587). For scale, Claude Opus 4.8 scored 1,562 and Gemini 3.1 Pro scored 1,558 on the same board. K3 ranked #1 in six of seven domains; its only second-place finish was in Gaming, behind Fable 5. That 17-place jump from Kimi K2.6 is real and it is large. Frontend code generation is the one axis where K3 is unambiguously the best model on the board right now, open or closed.

On the Artificial Analysis Intelligence Index, K3 scores 57, which places it around #3 to #4, on par with GPT-5.5 and behind only Claude Fable 5 and Claude Opus 4.8. The exact rank wobbles depending on how you count model variants: Artificial Analysis' own writeup says #3, some trackers show #4 once GPT-5.5's max and xhigh reasoning tiers are counted as separate entries, and a Kimi Code CLI run scored the same 57 but landed at #5. The score is solid and verifiable. The precise rank is not, so do not anchor a decision on "#3 versus #4."

Now the caveat that should govern how you read every coding number below. The three vendors do not grade on the same exam. DeepSeek V4 Pro reports 80.6% on SWE-bench Verified. GLM-5.2 reports 62.1 on SWE-bench Pro. Moonshot reports K3 through its own unified harness. Different benchmarks, different testing harnesses, no shared control. You cannot line these three coding scores up in a row and read a winner off the biggest number, and any comparison article that does is selling you a false precision. This is exactly the failure mode we flagged when DeepSeek and Qwen first cracked the open-source frontier: the benchmark a vendor chooses to report is itself a marketing decision.

Here is the three-way comparison with every figure traced to a source. Read the license and "weights" columns before the benchmark columns, because they change what the benchmarks even mean.

A few things jump out. All three ship a 1-million-token context window, so long-context capacity has stopped being a differentiator at this tier; it is table stakes now. DeepSeek V4 Pro carries the largest declared active-parameter footprint at 49B and the highest verifiable coding result. GLM-5.2 has the smallest total at 744B (one vendor tracker cites 753B; the head-to-head reporting uses 744B) and roughly 40B active, which makes it the lightest to self-host. K3 is the only multimodal model of the three; DeepSeek V4 Pro and GLM-5.2 are both text-only.

Architecturally, DeepSeek V4 Pro is worth a closer look if you care about the internals: its MoE layers use 384 routed experts plus one shared expert with six active per token, and it combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) in a hybrid attention stack. That is a different sparsity philosophy from K3's 16-of-896 routing, and it is part of why DeepSeek serves so cheaply.

This is where the decision usually gets made, and where K3's headline lead runs into gravity.

Kimi K3's API is $3.00 per million input tokens on a cache miss and $15.00 per million output tokens, with cached input discounted 90% to $0.30. Against its own predecessor Kimi K2.6 ($0.95 input / $4.00 output), K3 is roughly 3.2x more expensive on input and about 3.75x more expensive on output. So the quality jump came with a real price jump, not a free lunch.

Now put it next to the other two on the metric that actually reflects mixed real traffic, the Artificial Analysis blended price:

Read that gap carefully. DeepSeek V4 Pro serves at roughly $0.18 per million blended tokens, more than ten times cheaper than K3's $2.31, while posting 80.6% on SWE-bench Verified. For any high-volume agentic workload, coding loops, batch code review, classification, retrieval synthesis, that 10x is not a rounding error. It is the difference between a $2,000 monthly bill and a $23,000 one at the same token volume.

Across the model migrations we have run at Particula Tech, the most common and most expensive mistake is teams anchoring on the headline leaderboard rank and quietly eating a 5x to 10x serving premium on traffic where the cheaper model would have cleared the bar. The blended-cost column is the one that shows up on the invoice, and it is the one most evaluation decks bury on slide 40. K3 earns its premium only where its specific edge, top-tier frontend generation, multimodal input, the last few points of agentic reliability, directly moves your product. Everywhere else, the premium is pure margin leak.

Here is the fact that should govern any "which open model to run" decision this week: two of these three are open in the sense that matters, and one is open in the sense of a press release.

DeepSeek V4 Pro and GLM-5.2 both have full weights on Hugging Face today, MIT-licensed, unrestricted for commercial use, fine-tuning, and self-hosting. You can pull them, run them on your own GPUs, reproduce the benchmarks, audit the behavior, and pin a version that no vendor can change under you. Their self-host claims are checkable right now.

Kimi K3's weights are only committed for July 27, 2026. That date is in the future relative to this article. Until it arrives, K3's open-weight claims cannot be independently verified, because there is nothing to independently run. Every K3 number in this post came off Moonshot's hosted endpoint and Moonshot's harness. That does not make them wrong. It makes them unaudited, and "unaudited vendor benchmark" is a category every practitioner should hold at arm's length until the weights confirm it.

Assume the July 27 drop lands on schedule. What does open-weight actually buy you for a model this size? Less than the marketing implies, at first. A 2.8-trillion-parameter MoE has to fit its entire parameter set into cluster memory even though only 16 of 896 experts fire per token, because you cannot predict which experts a given token will route to. That is a multi-node, high-VRAM deployment, materially heavier than serving DeepSeek V4 Pro's 49B active path or GLM-5.2's ~40B active path. For most teams the honest near-term answer is that K3's open weights buy auditability and a self-host option you will not exercise until the tooling and quantized builds catch up. If you want the concrete hardware and quantization playbook for the lightest of the three, our guide to running GLM-5.2 locally walks the vLLM and GPU setup end to end, and the same shape of reasoning applies when K3's weights arrive.

Strip it down to the decision you are actually making this week.

If you want the best coding quality on a hosted API and open weights are a nice-to-have, not a gate: use Kimi K3. It is the #1 frontend coding model on a public board, it sits in the Opus 4.8 and GPT-5.5 tier on the Intelligence Index, and it is multimodal. You are paying roughly $2.31 blended and trusting Moonshot's harness until July 27. That is a defensible trade for a product where frontend or agentic quality is the differentiator and volume is moderate.

If you need open weights you can download and audit right now: use DeepSeek V4 Pro or GLM-5.2. Both are on Hugging Face today, both are MIT, both are runnable and checkable this afternoon. K3 is not an option in this lane yet, full stop, and pretending otherwise until July 27 is how teams end up blocked on a launch date they do not control.

If serving cost dominates: use DeepSeek V4 Pro. At ~$0.18 blended with 80.6% on SWE-bench Verified, it is the cheapest credible path to frontier-adjacent coding, hosted or self-hosted. The break-even between paying its API and self-hosting the weights is a real calculation, not a vibe, and we work it out in detail in the self-host versus API break-even math for 2026.

My opinionated default for most teams, today: run DeepSeek V4 Pro as the workhorse because it is cheap, verifiable, and downloadable; keep GLM-5.2 as the lighter self-host option when footprint matters; and pilot Kimi K3 on the hosted API for the specific frontend and agentic surfaces where its lead is real. Then re-run the whole evaluation on July 27 once K3's weights are public and its numbers can finally be checked. Do not commit self-hosted infrastructure to a model whose weights do not exist yet.

So which open model for coding 2026 fits your case? Here is the single lookup, and every cell traces to a verified figure above.