Mistral AI Goes Small with the "World's Best Edge Models," Ministral 3B and Ministral 8B
A year on from Mistral 7B's release, Mistral AI's latest models target privacy-centric on-device inference.
Artificial intelligence specialist Mistral AI has announced what it claims are "the world's best edge models," one year after the release of its seven-billion-parameter Mistral 7B: "les Ministraux," Ministral 3B and Ministral 8B.
"These models set a new frontier in knowledge, commonsense, reasoning, function-calling, and efficiency in the sub-10B [parameter] category, and can be used or tuned to a variety of uses, from orchestrating agentic workflows to creating specialist task workers," the company claims of its latest compact models. "Both models support up to 128k context length (currently 32k on vLLM) and Ministral 8B has a special interleaved sliding-window attention pattern for faster and memory-efficient inference."
Mistral AI is positioning the two new models for on-device inference in workloads such as translation, analytics, voice assistants, and even autonomous robotics, though it also promotes their use in partnership with its larger models for "multi-step agentic workflows" where some inference happens on-device with the rest farmed out to cloud systems as required.
According to the company's internal testing, the Ministral models beat equivalent models including Google's Gemma 2 2B and Meta's Llama 3.2 3B in metrics including multilingual operation, math, coding, and "knowledge and commonsense" — the latter being the current Holy Grail of AI model research, as companies wrestle with improving user trust in the face of continued issues with hallucinations and other artifacts of the technology. Moreover, the company even claims that the Ministral 3B model can outperform its own Mistral 7B "on most benchmarks."
The Ministral 8B model, the larger but more capable of the two, is being made available under the Mistral Commercial license for self-deployed use and at a $0.10-per-million-tokens price for use on the company's own cloud platform, with model weights available to researchers under the Mistral Research license; the more compact Ministral 3B is also available under the Mistral Commercial License and on the company's cloud platform at $0.04-per-million tokens, but the model weights are not available to researchers. "Both models will be available from our cloud partners shortly," the company promises.
More information on the models is available on the Mistral website.