DeepSeek V4 Pro 0813 1.6T vs Kimi K3

Analysis

DeepSeek V4 Pro 0813 and Kimi K3 are the two largest frontier open-weight releases tracked on this site as of August 2026, from rival Chinese labs six days apart (DeepSeek's version bump landed August 12, Kimi K3's weights opened July 27). Kimi K3 is unambiguously the bigger model, 75% more total parameters and more than double the active count, but on the one benchmark both self-report using the same suite name, Terminal-Bench 2.1, the two land within half a point of each other. Independent measurement tells a different story.

DeepSeek V4 Pro 0813 is a 1.6T-parameter MoE with 49B active per token, spread across 61 layers that all use the same hybrid Compressed Sparse Attention and Heavily Compressed Attention design, compressing every layer's KV cache by roughly 10x. Kimi K3 takes a sparser route on layers themselves: of its 93 layers, only 24 accumulate a KV cache at all (gated MLA), and the other 69 run Kimi Delta Attention, a linear-attention design whose memory is a fixed-size state that doesn't grow with context. At the 1,048,576-token context both models cap out at, an identical native window, DeepSeek's KV cache is 13.1 GB against Kimi K3's 51.5 GB, nearly 4x larger despite K3 caching barely a quarter of its layers (24 of 93) against DeepSeek's full stack (61 of 61). Compression beats sparsity here: DeepSeek's 10x per-layer squeeze on every caching layer outweighs Kimi K3 simply having fewer layers that cache in the first place. Weights tell a more straightforward story: both ship natively in low-bit formats (DeepSeek's routed experts in FP4/FP8 with no lossy community quant to fall back on yet, Kimi K3's in MXFP4 with MXFP8 activations), so the generic quantization ladder this site otherwise uses is a rough placeholder for both, not a real downloadable file. Using each model's own recommended-quant estimate anyway, DeepSeek needs about 609.6 GB of weights, Kimi K3 about 1,400 GB, and the full picture at the shared million-token context is 697.4 GB for DeepSeek against 1,625.7 GB for Kimi K3, more than double. On the one shared benchmark field both report, Terminal-Bench 2.1, the two nearly tie: DeepSeek's own comparison run scores 87.9, Kimi K3's own KimiCode-harness run scores 88.3, a 0.4-point gap that's well within noise given they're different labs' own measurements on different harnesses. Independent scoring from Artificial Analysis disagrees more clearly: its Intelligence Index, a nine-benchmark composite, puts Kimi K3 at 60 against DeepSeek V4 Pro 0813's 53, and DeepSeek hasn't republished GPQA Diamond, SWE-bench, or MMLU-Pro specifically for the 0813 checkpoint, so a wider academic comparison isn't possible from vendor-reported numbers alone; Kimi K3 reports GPQA Diamond 93.5 and SWE-bench Verified 76.8 with nothing directly comparable on DeepSeek's side. Licensing favors DeepSeek outright: MIT with no carve-outs, against Kimi K3's custom Kimi K3 License, MIT-like but requiring a separate agreement for AI-service businesses over $20M in any 12-month period, and mandatory "Kimi K3" branding for products over 100M monthly active users. Modality favors Kimi K3: it accepts image and video input through a 401M-parameter MoonViT-V2 encoder, where DeepSeek V4 Pro 0813 is text-only despite the April preview's own record on this site claiming vision support. On Artificial Analysis's own hosted-API measurements, not this site's local-hardware estimates, since neither model is remotely local-hardware territory, DeepSeek runs cheaper and faster: roughly $0.18 against Kimi K3's $2.31 per million tokens on AA's blended pricing measure, 83 tokens/second output against 41, and 1.63 seconds to first token against 3.27.

Bottom line: Neither model is a realistic self-hosting candidate: DeepSeek's recommended build alone needs more weight-only VRAM than the largest single machine this site tracks, a 512 GB Mac Studio M3 Ultra, and Kimi K3 needs roughly 2.3x that. For the API-only reality both models actually live in, the choice comes down to what you're optimizing for. Kimi K3 is the more capable model by Artificial Analysis's independent measure and the only one of the two with image and video input, but costs about 13x more per token and runs at roughly half the throughput on AA's own numbers. DeepSeek V4 Pro 0813 is the pragmatic default: near-identical Terminal-Bench 2.1 performance by each vendor's own count, a cleaner MIT license, and a fraction of the price. Auditability, fine-tuning, or an air-gapped requirement are the only real reasons to look at either model's weights directly, and even then DeepSeek is the less extreme ask of the two.