Is DeepSeek V4 or Kimi K3 the better open-weight model in 2026?

Kimi K3 wins on raw intelligence and adds native vision — it scores ~57 on the Artificial Analysis Intelligence Index versus ~44 for DeepSeek V4 Pro at max reasoning. DeepSeek V4 wins on cost: V4 Flash is $0.14/$0.28 per MTok (off-peak), the cheapest frontier-class API, and V4 Pro is $0.435/$0.87. Pick K3 for multimodal and top scores; pick DeepSeek V4 for the lowest bill.

How much do DeepSeek V4 and Kimi K3 cost via API?

Kimi K3 is $3 per million input tokens and $15 per million output (flat, no peak surcharge). DeepSeek V4 Pro is $0.435/$0.87 off-peak (2x during peak UTC windows), and V4 Flash is $0.14/$0.28. DeepSeek is roughly 7-20x cheaper per token, but Kimi K3 scores higher and handles images and video.

Can I self-host DeepSeek V4 and Kimi K3?

Both are open-weight under permissive (MIT / modified-MIT) licenses. DeepSeek V4 Flash (284B total, ~13B active) runs on a single 80GB GPU when quantized. Kimi K3 is a 2.8T-parameter MoE — Moonshot recommends 64+ accelerators (≈8 servers of H100/B200) for production, so it is far heavier to self-host.

Quick Answer

DeepSeek V4 vs Kimi K3: Open-Weight Showdown 2026

Published: July 28, 2026

The Short Answer

Kimi K3 is the stronger open-weight model; DeepSeek V4 is the cheaper one. As of late July 2026, Kimi K3 (open weights released July 27, 2026) scores about 57 on the Artificial Analysis Intelligence Index — #1 among open models and ahead of Claude Opus 4.8 — while DeepSeek V4 Pro sits around 44 at max reasoning effort. But DeepSeek V4 undercuts it dramatically on price and is trivially easier to self-host at the Flash tier.

Head-to-Head

	Kimi K3	DeepSeek V4 Pro	DeepSeek V4 Flash
Released	Jul 16, 2026 (weights Jul 27)	GA mid-Jul 2026	GA mid-Jul 2026
Params (total)	~2.8T MoE	~1.6T MoE	284B MoE
Context	1M tokens	1M tokens	1M tokens
Vision	✅ text+image+video	❌ text-only	❌ text-only
API in/out ($/MTok)	$3 / $15 (flat)	$0.435 / $0.87 (off-peak)*	$0.14 / $0.28 (off-peak)*
Intelligence Index	~57	~44	lower
License	modified MIT	MIT	MIT

*DeepSeek applies a 2x peak-hours surcharge (roughly 01:00–04:00 and 06:00–10:00 UTC). The legacy deepseek-chat / deepseek-reasoner aliases were retired July 24, 2026 — migrate to the new V4 model IDs.

Where Each One Wins

Pick Kimi K3 if you need the strongest open-weight intelligence, native multimodal input (images/video), always-on thinking, or a #1 finish on the Frontend Code Arena. It’s the first open model to pass FrontierCode 1.1.
Pick DeepSeek V4 Flash for high-volume pipelines — classification, extraction, summarization — where cost per token is the deciding factor. It is the cheapest actively-supported frontier-class API in July 2026.
Pick DeepSeek V4 Pro as a value coding/reasoning workhorse: ~80.6% SWE-bench Verified at a fraction of closed-model output pricing.

Self-Hosting Reality Check

DeepSeek V4 Flash quantizes down to a single 80GB GPU, making it the practical choice for teams that want to own the stack. Kimi K3’s 2.8T MoE is a data-center-scale deploy: Moonshot points to supernode configs of 64+ accelerators, so most users will consume K3 via API rather than self-host it.

Sources

Moonshot AI — Kimi K3 open weights (July 27, 2026): github.com/MoonshotAI/Kimi-K3
DeepSeek — V4 GA and pricing: deepseek.ai/pricing
MarkTechPost — Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2 comparison (July 18, 2026): marktechpost.com
Artificial Analysis — Intelligence Index: artificialanalysis.ai