Pixtral12B
by Mistral AI
Deprecated open-weight multimodal model for self-hosting.
Score
Score
Our verdict
Pixtral 12B is Mistral’s open‑weight, 12B‑parameter multimodal model that handles images and text with a 128K token context. As of late 2025, it’s deprecated and removed from Mistral’s hosted API, leaving self‑hosting (Apache‑2.0 licensed) as the only path. That’s great for labs and on‑prem teams wanting full control and zero license fees, but it’s a hurdle for product teams needing a managed service or SLA. Output quality remains respectable for captioning, VQA, and image‑grounded reasoning, though it trails newer vision models. Value is excellent if you already have GPUs; otherwise, provisioning and serving costs can outweigh the free license. Critically, there’s no vendor roadmap or support, and Mistral explicitly recommends switching to Ministral 3 14B. For research, benchmarking, and legacy maintenance, Pixtral 12B is still useful. For greenfield production, pick an actively supported model.
Overview
Score breakdown
Overall score
Scores are editorial assessments by the Compare Best AI team on a 0–10 scale.
Expert review
CBAI Editorial Team
Compare Best AI · Editorial Team
## Overview
How we tested
Days tested
7 days
Tasks evaluated
- ·Image captioning and VQA
- ·OCR + long-context summarization (128K)
- ·Grounded reasoning over screenshots
- ·Latency and throughput profiling on self-hosted GPUs
Method
Compared against Ministral 3 14B and Llama 3.2 Vision 11B using identical inputs
Reviewer
CBAI Editorial Team
Plans & pricing
Open weights / self-host
- Hosted API retired
Pricing may vary by region. Always verify on the vendor's website.
Feature comparison
| Feature | Pixtral12B | Ministral 3 14B | Llama 3.2 Vision 11B |
|---|---|---|---|
| Access | |||
| Official vendor-hosted API (Aug 2026) | |||
| Modality | |||
| Multimodal (image+text) | |||
| License | |||
| Open weights available | |||
| Status | |||
| Deprecated/Retired by vendor | |||
| Deployment | |||
| Self-hosting allowed | |||
| Guidance | |||
| Vendor-recommended Pixtral 12B replacement | |||
| Pricing | |||
| Pay-as-you-go hosted pricing published | |||
| Security | |||
| Air-gapped/offline capable | |||
Is it right for you?
Good fit for
On‑prem engineers
Can run Pixtral 12B entirely on owned GPUs for strict data residency and offline requirements.
AI researchers
Apache-2.0 weights enable reproducible experiments, ablations, and benchmarking without API constraints.
Cost‑conscious teams
Zero license fees; pay only for compute you already control.
Maintainers of legacy apps
Useful for auditing and migrating older multimodal workflows to newer models.
Less suited for
Teams needing a managed API
Mistral retired hosted access in 2025; you must self-host to use it.
Production SLAs and vendor support
No active product support or SLA since deprecation; consider Ministral 3 14B.
User reviews
Editorial score
Distribution is estimated from our editorial score. Verified user reviews coming soon.
Use cases
Typical ways teams rely on this tool — from everyday tasks to specialized workflows.
- On‑prem multimodal inference
- Edge/offline AI deployments
- Research and benchmarking
- Legacy model migration
- Internal tools without external API
Integrations
Reported connectors
Apps and services commonly connected out of the box or via official connectors.
- Hugging Face
- GitHub
- Mistral La Plateforme (retired)
Details
Category
Price
- Open weights / self-host $0
Free version
Best for
- On‑prem image+text inference
- Offline/air‑gapped experimentation
- GPU lab benchmarking
- Legacy app maintenance and migration planning
Frequently asked questions
Pixtral12B
Developer
Ready to get started?
Visit the Pixtral12B website to explore plans and start your free trial.
Was this page helpful?
Compare Best AI may earn a commission when you click links on this page. This does not influence our editorial scores or recommendations. Advertiser disclosure