Pixtral vs InternVL3.5
Pick InternVL3.5 for any new self-hosted OCR or document question answering project, because it is actively offered in nine sizes under Apache-2.0. Pick Pixtral only if you already call pixtral-12b-2409 on the Mistral API and want to keep that integration running. Pixtral is marked deprecated by Mistral, so it is a poor base for new work.
Pixtralmistral/pixtral | InternVL3.5opengvlab/internvl | |
|---|---|---|
| Type | Vision model | Vision model |
| Price | 2 credits / page | 2 credits / page |
| Per 1,000 pages | $2.00 | $2.00 |
| Typical speed | 3.0 s / page | 3.0 s / page |
| Best for | Questions about document images | Self-hosted document understanding |
| Inputs | images | images |
| Languages | Multilingual | Multilingual |
| Max pages per request | 500 | 500 |
| Tables | No | ✓ Yes |
| Handwriting | No | No |
| Math | No | ✓ Yes |
| Bounding boxes | No | No |
| Confidence | No | No |
| Forms | No | No |
| Multilingual | ✓ Yes | ✓ Yes |
| On OCRRouter | Available | Available |
Prices are what you pay on OCRRouter. Speed is the provider's typical figure, not a benchmark we ran.
Which to choose
Choose Pixtral if
- You want a hosted API: Pixtral was offered as pixtral-12b-2409 at $0.15 per million input and output tokens at launch.
- You need question answering over charts and documents, where Mistral reports 90.7 ANLS on DocVQA and 81.8 on ChartQA (CoT).
- You already have an integration on the Mistral API and its 128K token context lets you send many images per request.
Choose InternVL3.5 if
- You want to choose a size: there are nine, from 1B to 241B-A28B, so you can trade accuracy for cost.
- You have modest GPU hardware: models up to 30B run on a single A100, and the 20B-A4B and 30B-A3B MoE variants activate few parameters per token.
- You are starting a new project and want a family that is not marked deprecated, under Apache-2.0 and built on Apache-2.0 Qwen3.
Key differences
| Pixtral | InternVL3.5 | |
|---|---|---|
| How you run it | Mistral API (pixtral-12b-2409) or self-hosted open weights | Self-host only, with Transformers, LMDeploy or vLLM |
| Licence | Apache-2.0 | Apache-2.0 |
| Price | $0.15 per million input and output tokens at launch on the API; free to self-host, you pay for compute | Free, self-hosted, you pay for compute |
| Model size | One model, 12.4B (12B decoder + 400M vision encoder) | 1B to 241B-A28B; the 8B variant is 8.5B |
| Published OCR-related scores | DocVQA 90.7 ANLS, ChartQA 81.8 (CoT), MMMU 52.5 (CoT), all from Mistral | Model card shows OCR scores only as chart images, not text values |
| Hardware | Not published | Up to 30B on one A100; 38B needs two A100s; 241B-A28B needs eight A100s |
| Known failure | Reported drops on handwriting, small fonts and non-Latin scripts; Q4 quantization degrades small text; one HN user said OCR is not always close to correct | General VLM with no bounding boxes or layout output by default, so no word coordinates |
Sources
- Mistral announcement
- Mistral Sept 2024 pricing
- HN launch thread
- Third-party deployment guide
- Hugging Face model card (8B)
- InternVL GitHub
What each one is
Pixtral
Mistral's open-weight vision models, Pixtral 12B and Pixtral Large. Mistral's API no longer serves them, so they run self-hosted.
Pixtral detailsInternVL3.5
Shanghai AI Lab's open vision-language family (1B to 241B, August 2025), strong on OCR and document benchmarks.
InternVL3.5 detailsTry both on your own file
Run the file once with each model and compare the output.
Use one, fall back to the other
{"model": "mistral/pixtral","fallbacks": ["opengvlab/internvl"],"file_url": "https://example.com/document.pdf"}Questions
More comparisons
Pixtral alternativesInternVL3.5 alternativesBest OCR models in 2026