AI agents · OpenClaw · self-hosting · automation

Quick Answer

Best Open-Weight AI Models 2026: Ranked by Value

Published:

Best Open-Weight AI Models 2026: Ranked by Value

Open weights crossed the frontier in 2026: Kimi K3 now ranks #3 overall, DeepSeek serves near-flagship coding for pennies, and permissive licenses make on-prem real. Here are the best open models ranked by capability and cost.

Last verified: July 25, 2026

Quick Rankings

RankModelLicenseInput / OutputCost / taskBest for
1Kimi K3Open (Jul 27)$3 / $15~$0.165Peak capability + vision
2DeepSeek V4 ProMIT$0.435 / $0.87*~$0.017Cheapest frontier-ish
3GLM-5.2Open~low~lowBalanced self-host
4MiniMax M3Open$0.60 / $2.40~$0.0301M context, coding
5DeepSeek V4 FlashMIT$0.14 / $0.28*~$0.006Highest-volume, cheapest

DeepSeek prices are off-peak; they double during peak UTC hours (1-4 and 6-10).

1. Kimi K3 — the capability leader

Moonshot AI’s Kimi K3 (2.8T MoE, 1M context, native vision) scores ~57 on the Artificial Analysis Intelligence Index, ranking #3 overall — the first open-weight model in the top tier with GPT-5.6 Sol and Claude Fable 5. Open weights land July 27, 2026. At flat $3/$15, it’s about half the cost of US flagships for comparable intelligence. Best for difficult reasoning, judgment, and multimodal work where open weights matter.

2. DeepSeek V4 Pro — cheapest near-frontier

MIT-licensed and cheap: at $0.435/$0.87 off-peak (~$0.017/task), DeepSeek V4 Pro (1.6T total / 49B active) scores 44 on the intelligence index but leads open weights on coding — DeepSeek V4 Pro-Max hits ~80.6% SWE-bench Verified, tied with Gemini 3.1 Pro. Cache hits drop to ~$0.043/MTok. Best for high-volume text and coding at rock-bottom cost.

3. GLM-5.2 — balanced self-host

GLM-5.2 scores ~51 on the intelligence index — between DeepSeek V4 Pro and Kimi K3 — and is a popular, license-friendly choice for teams self-hosting a capable generalist without K3’s infrastructure demands. Best for balanced on-prem deployments.

4. MiniMax M3 — long context, coding

MiniMax M3 offers a 1M context and 80.5% SWE-bench Verified at $0.60/$2.40 ($0.030/task) — a strong middle option when you need long context and solid coding cheaply. Best for long-document coding workflows.

5. DeepSeek V4 Flash — cheapest at any scale

At $0.14/$0.28 off-peak (~$0.006/task, cache hits ~$0.0028), DeepSeek V4 Flash is the cheapest capable open model for extraction, classification, and high-volume chat. Note: the legacy deepseek-chat / deepseek-reasoner API names were retired July 24, 2026. Best for maximum volume at minimum cost.

How to Choose

  • Need peak capability or vision? Kimi K3.
  • Need cheapest near-frontier coding? DeepSeek V4 Pro (MIT).
  • Balanced self-host? GLM-5.2.
  • Long context + coding? MiniMax M3.
  • Cheapest at any scale? DeepSeek V4 Flash.

Bottom Line

Open weights are no longer the “budget fallback.” Kimi K3 sits inside the frontier, DeepSeek delivers ~80% SWE-bench coding for pennies, and permissive licenses make all of these viable as primary, self-hostable models in 2026.

Sources