Mistral — Dense + Pixtral VL
Mistral dense transformers and Pixtral vision-language models. GGUF through all GPU backends; NPU via a sliding-window-attention (SWA) generator.
Models
| Model | Params | 1BP Size | Backend(s) | Perf |
|---|---|---|---|---|
| Mistral-7B-v0.3 | 7B | 4.3 GB | GGML-Vulkan / ZINC / NPU / HIP | — |
| Ministral-8B | 8B | 4.7 GB | GGML-Vulkan / ZINC / NPU / HIP | — |
| Pixtral (VL) | 12B | — | GPU HIP (vision) | 🔄 |
Notes
- NPU: Mistral-7B Peano build stanza (
build_mistral_7b) — GEMM dims match Llama-3.1-8B. SWA attention MLIR generator (generators/n1_core_swa.py) handles sliding-window attention. Build with./build_xclbins.sh mistral_7b. - GPU HIP: GGUF validated. Vulkan / CPU: functional, perf pending.
See also: full model support detail · benchmarks SSOT · all families