The open weights leaderboard · for developers

Best for open weights

Decision area
Best models you can self-host today.
Editor picks
7 editor picks
Eligible models
143 eligible models
See raw /best
EDITOR'S CHOICEResearched 34d ago

DeepSeek V4 Pro

DeepSeek · 1M context
Excellent

Models with downloadable weights, chosen on hosting options, license and price.

LiveCodeBench (93.5), 80.6 SWE-bench, 1M context, $0.87 out.

The numbers
$/1M out
$0.73
$0.36 input
Context
1M
max window
Pros
  • +Frontier-class coding, open weights
  • +1M context
  • +$0.87 / 1M out
Cons
  • −Heavy to self-host at full size

Also worth picking

The runners-up

ranked by editorial pick orderEditorial tiersExcellentStrongSolid
Moonshot AI · 1M
$15.00 / 1M out
Moonshot AI's 2.8T-parameter MoE with weights on Hugging Face under the Kimi K3 license. It has a 1M window, text, image and video input, and tool calls, at $3 / $15 per 1M tokens hosted.
AI at Meta · 1M
$4.25 / 1M out
Meta's 816B MoE (about 42B active) for coding and agentic work, released October 15, 2026 under the Muse Model License, which has conditions. It takes text, image, video and audio input. Hosted at $1.25 / $4.25 per 1M tokens on Together AI and Fireworks.
DeepSeek · 1M
$1.20 / 1M out
552B MoE under the MIT license, with a 1M window, up to 384K output tokens and vision input. Weights on Hugging Face.
Zhipu AI · 200K
$4.40 / 1M out
SWE-bench Pro 58.4 with excellent agentic behavior; an easy self-hosted default.
Alibaba · 256K
$3.00 / 1M out
Strong reasoning + multilingual baseline; widely hosted MoE that's cheap to serve.
NVIDIA AI · 1M
$0 / 1M out
NVIDIA's 120B MoE (about 12B active) with vision, under OpenMDW-1.1, with a 1M window. NVIDIA reports it reads text in images, video, web pages and PDFs.

Eligibility

143 models are eligible for this board

Eligibility means tagged with useCases: [open-weights]. Pins must come from this pool.
All picks