Loading the catalogue…
Loading the catalogue…
A 12B open-weights instruction model co-developed by Mistral AI and NVIDIA, featuring a 131k-token context window, the efficient Tekken tokenizer, native tool calling, and multilingual support across nine languages — available under Apache 2.0 for unrestricted self-hosted inference.
Each axis is the mean score across the family’s variants that have been scored on that dimension. Per-axis sample size is shown next to each label — the family currently aggregates up to 1 variant per axis.
Values aggregated across the family’s variants: any variant supporting a capability resolves the family to Supported; flag-driven support resolves to Optional; only when every variant explicitly denies a capability does the family render as Not supported. 15 of 15 capabilities have variant data so far.