Mistral Large 4 weights & Hugging Face guide
Track the planned weights release, check the model card and estimate download and deployment needs.
On this page
Release status
The official October 6 announcement schedules open weights for the end of October. That schedule is not evidence that files are already available. Check the official announcement and Mistral AI’s Hugging Face organization for the actual model repository.
Identify the right repository
Verify the publisher, exact model identifier and revision. A repository name alone does not establish that it is official. Read the model card, supported architecture and file list. Pin a commit or revision instead of deploying a moving branch.
Before you download
Check the license in the released repository; do not assume a license from an earlier model. Inspect file checksums and expected sizes. Calculate disk capacity for the original weights, conversion artifacts, cache and rollback copy. Never execute scripts from an unverified mirror.
Memory estimate
A 1.05T-parameter model requires roughly 2.1 TB at two bytes per parameter, 1.05 TB at one byte or 525 GB at half a byte. These are weight-only arithmetic estimates, not tested GPU configurations. Quantization metadata, routing, KV cache and runtime overhead increase the total. Active parameters do not reduce the need to store the full expert weights.
Framework readiness
Confirm that your runtime supports the exact architecture and quantization format. Test tokenizer compatibility, expert placement, image processing and multi-device communication. Use a short prompt and known image before attempting long context.
Deployment checklist
- Record the repository revision and license.
- Verify downloaded files before conversion.
- Measure memory, latency and throughput with realistic inputs.
- Set request limits and cancellation behavior.
- Keep credentials outside public code and log only necessary diagnostics.
If weights are not available
Use the official API for experiments if you have access. This workspace does not distribute weights and has not benchmarked a self-hosted Large 4 deployment. Read our local deployment guide before budgeting hardware.