Mistral Small 4 Lands

Mistral published Small 4 on March 16, 2026 alongside updated docs and a model card that describe its unified deployment options and runtime knobs. (docs.mistral.ai) Mistral’s official documentation lists roughly 6 billion activated parameters per token (about 8B when counting embedding and output layers), while the model’s Hugging Face repo cites an activated‑per‑token figure closer to 6.5 billion, creating a small reporting discrepancy across sources. (mistral.ai) The public Hugging Face release includes multiple checkpoints and efficiency artifacts — an FP8 checkpoint, an NVFP4 quantized checkpoint, and a separate “eagle” speculative‑decoding head — all intended to reduce memory and latency during inference. (huggingface.co) Files for the model were uploaded to Hugging Face (community mirrors and GGUF ports followed), with third‑party trackers and writeups noting a multi‑hundred‑gigabyte footprint for the primary artifact on the hub. (letsdatascience.com) Mistral announced a concurrent enterprise offering called Forge for end‑to‑end in‑house model training and named early partners including ASML, Ericsson, the European Space Agency, DSO National Laboratories Singapore, HTX Singapore, and Reply. (mistral.ai) The company published explicit hardware guidance for on‑prem deployments, listing minimum and recommended multi‑GPU cluster configurations (examples include several HGX H100/H200 or DGX B200 nodes for production runs). (mistral.ai) Early community threads report integration work remains ongoing for some optimized checkpoints (notably NVFP4/GGUF variants) with vLLM and other inference backends due to nonstandard head sizes and compatibility issues flagged by developers. (forums.developer.nvidia.com)