mammoth-1 27B

A 27.0B-parameter model called mammoth-1 has been released under an MIT license with weights on HuggingFace, built for a single task: returning a confidence number per answer option in one forward pass for form-filling, support-ticket routing and classification rather than generating text. The model carries a 32k context window, originated in the USA, and was released in Oct 2026; its vendor-stated claim of ranking #1 of 208 on the Jev-index is not independently verified, though the model's probabilities can be checked in an afternoon. The release matters because routing and gating decisions at scale can be made without the cost of generating text for each judgment.

mammoth-1 is a model with one job: filling in forms. Hand it a support ticket and a question with a fixed set of answers, and it returns a confidence number per option in one forward pass. No paragraph is written, no tokens stream; think of a data-entry clerk who reads the whole document and stamps a box. It exists because routing, classification and gating need decisions at scale, and generating text to decide something is wasteful. Reach for it when every request needs a judgment and cost per judgment matters. Honest status: weights on HuggingFace MIT , the 1-of-208 Jev-index claim is vendor-stated; probabilities are verifiable in an afternoon. decision-model classification routing local - Parameters - 27.0B - Context - 32k - License - mit - Origin - 🇺🇸 USA - Released - Oct 2026 PRICE HISTORY Inference cost over time Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API. Loading price history...