Mistral announced the launch of Mistral Small 4, a 119‑billion‑parameter hybrid model that merges the capabilities of its flagship Magistral reasoning engine, Pixtral multimodal system, and Devstral coding agent into a single, versatile architecture. The model uses a mixture‑of‑experts design with 128 experts and 4 active per token, and it offers a 256k‑token context window to support long‑form interactions and document analysis. Mistral Small 4 is configurable through a reasoning_effort parameter that lets users toggle between fast, low‑latency responses and deep, step‑by‑step reasoning for complex problems. The model is available under an Apache 2.0 license, can be fine‑tuned, and is distributed on NVIDIA‑accelerated platforms such as vLLM, llama.cpp, SGLang, and Transformers. Performance benchmarks show that with reasoning enabled, Mistral Small 4 matches or surpasses GPT‑OSS 120B on three tests while producing significantly shorter outputs, which translates to lower latency and inference costs. The release also marks Mistral’s entry as a founding member of the NVIDIA Nemotron Coalition, emphasizing collaboration on open‑source AI development. The company highlights that the model’s efficiency and configurability make it suitable for developers building coding automation tools and for enterprises deploying general chat assistants, document understanding, and multimodal analysis. Mistral Small 4 is available day‑0 as an NVIDIA NIM, providing optimized, containerized inference out of the box, and can be customized with NVIDIA NeMo for domain‑specific fine‑tuning. Technical documentation and support for enterprise deployments are available through Mistral’s AI Governance Hub.
Introducing Mistral Small 4
Mistral has released Mistral Small 4, a hybrid model that unifies reasoning, multimodal, and coding capabilities into a single, configurable system. The 119B‑parameter model supports text and image inputs, offers a 256k context window, and is available under an Apache 2.0 license. It is optimized for speed and throughput, delivering up to 40% lower latency and three times more requests per second than its predecessor. Mistral Small 4 is open‑source, can be fine‑tuned, and is available on NVIDIA’s accelerated platforms and as a day‑0 NVIDIA NIM.
Why it matters
By combining instruct, reasoning, and multimodal functions into one model, Mistral Small 4 reduces the need for multiple specialized models, streamlining deployment and lowering operational costs for AI developers and enterprises.