Quick Facts
Developer
Mistral AI
License
Apache 2.0
Typical Sizes
3B, 8B, 14B (cloud and local tags)
Best For
Lightweight assistants, RAG, image understanding, structured extraction
Recommended Deployment
8B cloud tag for most teams; local tags for regulated or air-gapped environments
Why Choose This Model
Ministral 3 is a reasonable default when you need multilingual text and image handling without the memory footprint of a larger model. It supports tool calling and a 256K context window, which covers most document Q&A, summarization, and lightweight agent use cases. It is not built for repository-scale coding or heavy multi-step agentic reasoning -- evaluate a larger model if that is your primary workload.
Terminal.Glass Deployment
Two deployment paths are supported. The cloud tags
(ministral-3:3b-cloud,
ministral-3:8b-cloud,
ministral-3:14b-cloud)
run inference on Ollama Cloud while Terminal.Glass hosts the interface, accounts, and your
document index locally. This requires a signed-in Ollama account and sends prompts, images, and
retrieved context off your infrastructure. The local tags
(ministral-3:3b,
ministral-3:8b,
ministral-3:14b) run entirely
on hardware you provision, with no data leaving your network. Both paths are compatible with
Terminal.Glass's standard Open WebUI, Ollama, and OpenClaw setup.
Recommended Uses
- Document Q&A and RAG over internal knowledge bases
- Multilingual support and translation workflows
- Image understanding paired with text prompts
- Structured JSON extraction and lightweight tool-calling agents
- Internal drafting, summarization, and support bots
Not recommended as a primary engine for large-scale code refactoring or multi-agent orchestration at scale -- pair with a larger model for those cases.
Hardware Guidance
Cloud tags need no local GPU -- a small VM, NUC, or mini-PC running Terminal.Glass and Ollama is sufficient since inference happens on Ollama Cloud. Local tags require a GPU sized to the model: the 3B and 8B variants run comfortably on a single mid-range GPU, while the 14B variant needs more VRAM headroom, particularly with long-context or image inputs. Running any of these without a GPU falls back to CPU inference, which is workable for light testing but too slow for regular team use. If you want to avoid GPU selection and sizing entirely, use the cloud tags. If you need full control over hardware and data, provision a GPU instance and run the local tags.
Honest Guidance
- Regulated or confidential data: use local tags on hardware you control, not the cloud tags.
- Steady daily use: local deployment on owned hardware is often cheaper over time than recurring per-token cloud cost.
- Air-gapped requirements: cloud tags require network access to Ollama Cloud; only local tags run fully offline.
- Version pinning: local deployment gives you a fixed checkpoint; cloud tags can change without notice.
- Fast pilots and evaluation: cloud tags are the quicker path with no upfront hardware decision.
Learn More
For detailed licensing terms, model card specifics, and full deployment documentation, see the Ministral 3 reference page on NoCloudGPT.