When data must not leave your infrastructure, you need a model whose weights you can download and run yourself.
3 models meet the requirements, ranked by monthly cost for this workload.
| Model | Arena score | Class | Context | Images | Input $/1M | Output $/1M | Monthly |
|---|---|---|---|---|---|---|---|
| Llama 4 Maverick Lowest cost | 1,327 | Mid | 128K | ✓ | $0.19 | $0.65 | $27 |
| Qwen3.8-27B | 1,436 | Mid | 262K | ✓ | $0.15 | $1.88 | $42 |
| DeepSeek V4 Pro | 1,465 | Reasoning | 1M | — | $0.66 | $1.98 | $89 |
Licenses: DeepSeek V4 Pro under MIT, Qwen3.8-27B under Apache 2.0, Llama 4 Maverick under the Llama Community License, which has its own conditions. Read the license before commercial use. API prices for Llama and Qwen come from third-party hosts.
Arena score: LMArena (text, style control), from human votes, published Oct 28, 2026, CC BY 4.0.
Our side-by-side tests for this task are in progress. Until they are published, run 10 to 20 of your own real examples on two or three candidates before you choose.
Independent leaderboards: LMArena · Artificial Analysis
How this list is built: models in the classes suited to the task that meet its hard requirements (image input, context size, open weights), ranked by price. A lower price does not mean better quality.