Kimi K3.1 Pro vs Kimi K3

Analysis

Kimi K3.1 Pro is Kimi K3 with much more post-training, released by the same lab three months later with the same parameter count, the same 896-expert layout, and the same 1M-token window. Both open-weight checkpoints are 96 MXFP4 safetensors shards within half a gigabyte of each other, so this is the rare comparison where the hardware question has a one-word answer: identical. What changed is what the model does with it.

On paper the two are the same machine: 2.8T total parameters with 104B active per token, 93 layers of which 69 run Kimi Delta Attention (a linear-attention design with a fixed-size state) and 24 keep a per-token cache. The only architecture change is inside those 24 layers, which move from gated MLA to Gated DeepSeek Sparse Attention. A small indexer picks the 2,048 most relevant earlier tokens for each new one, so long prompts cost less compute, but every token's compressed entry is still stored, so the KV cache is the same: 51.5 GB for each at the full 1,048,576-token window. The weights are 1,561.3 GB for K3.1 Pro against 1,560.9 GB for K3, and at an 8k context the totals are 1,749.1 GB against 1,748.7 GB; both fit one eight-GPU node of 288 GB B300s and neither fits anything smaller this site tracks. The differences are in behavior. Moonshot's launch post reports K3.1 Pro at 61.2 on its long-horizon coding test against K3's 55.95 (shown rounded to 56.0 below), cuts the share of wrong claims on its own hard hallucination test from 23.4% to 8.3%, and reports gains of 7.2 and 11.6 points on internal financial modeling and research tasks. It also says it removed cyber-related training data for this open release, yet the shipped model still scores 81.0 on Moonshot's exploit benchmark against K3's 33.3. The headline benchmarks moved less: GPQA-Diamond is 94.2 against 93.5. Moonshot moved its coding headline from Terminal-Bench 2.1 (K3's 88.3) to the newer Terminal-Bench 4.0 (65.7), so the two models have no published coding score on the same suite. Two smaller differences: K3.1 Pro's card lists image input only where K3 also took video, and the license is renamed with the same two conditions, a separate agreement for model-as-a-service businesses past $20M in 12-month revenue and a visible model name for products past 100M monthly users. All of the capability figures are Moonshot's own; Artificial Analysis independently places K3.1 Pro at 46 on version 5.0.1 of its Intelligence Index.

Bottom line: If you already serve Kimi K3, switch: the hardware, cache and parallelism layout carry over unchanged, and you get fewer confidently wrong answers, which was K3's weakest point. If you rely on video input, check it before migrating, because K3.1 Pro's card only claims image input. Nobody should pick between these two on hardware grounds, since there is nothing to choose: both need about 1.75 TB of pooled accelerator memory, and both are API-first models for anyone without a Blackwell-class node.