The best open-weight AI models

Updated February 2026 · 5 models · ranked by open-weight, by price

Open-weight models can be self-hosted for data control and cost predictability. These come from providers shipping open or open-weight releases, ranked by hosted price.

#ModelInputOutputContext
1Llama 4 Scout
Meta
$0.20$0.40512K
2DeepSeek V4
DeepSeek
$0.40$1.20256K
3Codestral 2
Mistral
$0.50$1.50256K
4Llama 4 Maverick
Meta
$0.60$0.901M
5Mistral Large 3
Mistral
$2.00$6.00256K

1. Llama 4 Scout

Meta · $0.20 in / $0.40 out per 1M · 512K context · Aug 2025

Compact open-weight model for cheap, self-hosted inference and edge deployment.

2. DeepSeek V4

DeepSeek · $0.40 in / $1.20 out per 1M · 256K context · Jan 2026

Open-weight MoE model with strong reasoning and coding at a very low price.

3. Codestral 2

Mistral · $0.50 in / $1.50 out per 1M · 256K context · Nov 2025

Code-specialized model tuned for autocomplete, fill-in-the-middle, and refactors.

4. Llama 4 Maverick

Meta · $0.60 in / $0.90 out per 1M · 1M context · Aug 2025

Open-weight model you can self-host — strong coding with permissive licensing.

5. Mistral Large 3

Mistral · $2.00 in / $6.00 out per 1M · 256K context · Feb 2026

European frontier model with strong multilingual and function-calling support.

Pricing is illustrative and per 1M tokens. Not sure which to pick? Run the model finder for a recommendation, or browse the full directory.