Mistral Small 4 119B is an open-weights hybrid model from Mistral that collapses three previously separate model families — instruction following, the former Magistral reasoning line, and Devstral-style coding — into a single checkpoint, letting users toggle between fast reply and deeper reasoning on demand. Architecturally it is a Mixture-of-Experts design with 119 billion total parameters and roughly 6.5 billion active per token across 128 experts with 4 active per forward pass, paired with multimodal inputs that accept both text and images while emitting text. Mistral publishes auxiliary artifacts alongside the main model, including a trained Eagle speculative-decoding head and an NVFP4 4-bit float quantization checkpoint, so the same weights can be deployed for either quick interactive use or higher-throughput batched serving.

In its own benchmark notes, Mistral reports a roughly 40 percent reduction in end-to-end completion time under a latency-optimized serving setup and about three times the requests-per-second of the prior Mistral Small 3 generation when tuned for throughput, framing the release as a generational efficiency jump rather than a wholesale capability reset. Independent activity has followed quickly: an NVIDIA developer-forum user published a detailed SGLang deployment and benchmark write-up for the model on a DGX Spark system within days of release, giving practical signal that the weights run on high-end consumer and workstation hardware. The practical fit is a general-purpose model that can switch into a reasoning mode for harder analytical or coding tasks, accept visual context alongside text, and be quantized or speculative-decoded to suit latency or throughput targets.