Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2: license, serving cost, benchmarks

Moonshot released Kimi K3 on July 16 and the open-weight race stopped being a two-horse story. K3 is a 2.8 trillion parameter MoE with 896 experts and 16 active per token, Kimi Delta Attention, native vision and a 1M context. Arena's blind test put it first on Frontend Code at 1,679 points, ahead of Fable 5, and Artificial Analysis ranks it fourth among all frontier models, above Claude Opus 4.8. The weights are due July 27 under a Modified MIT license. Two days later a comparison against DeepSeek V4 Pro and GLM-5.2 started circulating, and it is worth picking apart, because the benchmark numbers point one way and the deployment decision points another.

The three models on paper

• Kimi K3 (Moonshot): 2.8T total, 16 of 896 experts active, 1M context, 93.5% GPQA Diamond, 84.2% MCP Atlas. Moonshot's own coding numbers against GLM-5.2: DeepSWE 67.5 vs 46.2, FrontierSWE 81.2 vs 67.3. License: Modified MIT, weights not public until July 27.
• GLM-5.2 (Z.ai): 744B MoE, MIT, 1M context, released June 17. 91.2% GPQA Diamond, 62.1% SWE-bench Pro, Artificial Analysis Intelligence Index 51, which matches GPT-5.6 Luna at roughly $0.32 per task.
• DeepSeek V4 Pro: MIT, 1M context by default, 80.6% SWE-bench Verified at about a seventh of the output price of the frontier models it is compared against; the stable release lands July 24.

Two more names belong in the same bracket: MiniMax M3 at 59% SWE-bench Pro, and Qwen 3.6 under Apache 2.0. Both have the cleaner license story.

Why the license matters more than the delta

If you are building a paid tool, the question is not which model wins a coding benchmark by twenty points. It is whether you can ship the model to a customer's on-prem cluster, wrap it in a commercial product, and not have a legal review stall the deal. GLM-5.2 and DeepSeek V4 Pro run under plain MIT. K3 runs under a Modified MIT whose modifications you need to read before you promise anything to a client.

A benchmark point is a marketing asset; a license clause is a liability line.

The serving cost splits the same way. A 2.8T MoE is not download and run. Even with 16 experts active, the resident weights need sharding across serious infrastructure, and the bottleneck is memory, not compute. GLM-5.2 at 744B is still a big-iron model, but it is a different tier of hardware ask. DeepSeek V4 Pro gives you the lowest per-token API price of the three if you would rather not host at all, and DeepSeek's 2026 lineup already has a cheap Flash tier under it for routing.

What the coding gap actually tells you

Moonshot's 67.5 vs 46.2 on DeepSWE is a vendor-published comparison of its own model against a competitor. Take it as the upper bound of the gap. What is consistent across sources is the ordering: K3 sits above GLM-5.2 on agentic coding, and GLM-5.2 sits above the SWE-bench Pro 60% line that I treat as the threshold where building coding tooling over an open core beats paying for a frontier API. So both are viable; K3 is better and much more expensive to run.

The closed-side anchor for routing is Gemini 3.6 Flash at $1.50 in and $7.50 out; any self-host plan has to beat that on total cost including engineering time.

Build your own eval before you pick

• Take 30-50 real tasks from your own repo or workflow and run all three through the same harness.
• Measure tokens per completed task, not just pass rate; MoE models differ a lot in verbosity.
• Price the hosting honestly: GPU-hours or unified-memory boxes plus the engineer who keeps it alive.
• Check the license text against the way you will distribute the product, and get sign-off in writing.
• Keep an adapter layer so you can swap providers when the price, capacity or legal picture moves.

The honest gap

Every Kimi K3 number above is either Moonshot's own or from a leaderboard measured on Moonshot's hosted endpoint, because the weights are not public until July 27. Until they are, nobody has independently reproduced the coding gap on their own hardware, and the parameter count for GLM-5.2 is reported as 744B in one source and 753B in another. Treat the ranking as directional and the exact deltas as unverified until you have run them yourself.