Kimi K3 vs DeepSeek V4 Pro
Kimi K3 scores higher on 4 of the 5 benchmarks both models publish. The widest gap is on BrowseComp, where Kimi K3 leads by 7.8 points. Both have open weights; DeepSeek V4 Pro is far lighter to host, needing about 1.1 TB at 4-bit against 1.9 TB for Kimi K3.
Kimi K3Moonshot AI · OpenLLM Buddy catalog90.6composite
DeepSeek V4 ProDeepSeek · Open weights91.0compositeShared benchmarks
Raw published scores; the higher score on each row is highlighted. LMArena Elo is a human-preference vote and is not counted as a win.
- GPQA Diamond93.590.1
max reasoning effortV4-Pro (Preview) Max, pass@1
- Humanity's Last Exam43.542.7
HLE-Full, no tools, max effortV4-Pro-0813, no tools
- Humanity's Last Exam (with tools)56.060.0
HLE-Full, with tools, max effortV4-Pro-0813, with tools
- Terminal-Bench88.387.9
Terminal-Bench 2.1, max effortTerminal-Bench 2.1, V4-Pro-0813, max effort, DeepSeek Harness
- BrowseComp91.283.4
max effortV4-Pro (Preview) Max
- LMArena Text Elo · not counted14881465
LMArena text overall, 'Kimi K3 (Max)', ±4, as of 2026-10-27DeepSeek-V4-Pro (High, 20260813), ±7, board dated 2026-10-27
Should you pick Kimi K3 or DeepSeek V4 Pro?
Pick Kimi K3 if…
- You care most about GPQA Diamond, BrowseComp, where it scores at least 2 points higher.
- People prefer its answers in blind chats: LMArena Elo 1488 against 1465.
- You want it running in minutes on a dedicated GPU through OpenLLM Buddy, in an uncensored build.
Pick DeepSeek V4 Pro if…
- You care most about Humanity's Last Exam (with tools), where it scores at least 2 points higher.
- You want the lighter model to host: about 1.1 TB at 4-bit against 1.9 TB.
- You need fast responses: about 94 tokens per second through APIs, against 66.
- Cost matters: $3.96 per million output tokens against $15.00.
- You need a permissive license: it ships under MIT.
Hardware and cost to run
Kimi K3
Even at 4-bit it needs about 1.9 TB (Multiple nodes (over 1.1 TB)), well beyond a home workstation, so most people run it on rented multi-GPU servers. Full BF16 precision needs about 6.4 TB (Multiple nodes (over 1.1 TB)).
On OpenLLM Buddy: Kimi-K3-Uncensored-GGUF runs on 8× H100 80GB as a GGUF quantized build, with an OpenAI-compatible API and no per-token fees. Prices start at $422 USD for 11 hours.
Deploy Kimi K3DeepSeek V4 Pro
Even at 4-bit it needs about 1.1 TB (8× H200 (1,128 GB)), well beyond a home workstation, so most people run it on rented multi-GPU servers. Full BF16 precision needs about 3.7 TB (Multiple nodes (over 1.1 TB)).
Category profile
Category index uses every test each model reports, not only the shared ones.
Specs side by side
| Kimi K3 | DeepSeek V4 Pro | |
|---|---|---|
| Lab | Moonshot AI | DeepSeek |
| Weights | Open | Open |
| Released | 2026-07-27 | 2026-08-13 |
| Parameters | 2.8T total, 104B active (MoE) | 1.6T total, 49B active (MoE) |
| Context | 1M | 1M |
| License | Kimi K3 License (custom) | MIT |
| 4-bit memory | 1.9 TB · Multiple nodes (over 1.1 TB) | 1.1 TB · 8× H200 (1,128 GB) |
| API price | $3.00 in / $15.00 out per 1M tokens | $1.32 in / $3.96 out per 1M tokens |
| API speed | 66 tok/s | 94 tok/s |
Related comparisons
Sources
Scores are published results from model cards, lab reports, and public leaderboards, each linked to its source. Settings such as thinking mode or test version differ between sources and are shown next to each score. OpenLLM Buddy catalog models are uncensored fine-tunes; their scores are those of the official base model. Data collected 2026-10-30. How the index works