Model Details
K2-Type-0.9B
Quick answer: K2-Type-0.9B is a 0.9B-parameter decision model from IFM (Institute of Foundation Models), released September 26, 2026 under Apache 2.0. Built on the K2-Horizon-0.9B backbone with a pointer head, it takes a state and typed questions (
noul,choice,score) and returns a probability for every option in one forward pass, never generating text. It scores 76.2% (176/231) on the public JevBench set with 27 ms median latency on one H200.
At a Glance
What it is: A very small decision model in the style of TypeSafe AI's Jev that is wire-compatible with the /v1/systemone API.
Where it leads: Calibration and speed for its size — Brier 0.328, ECE 0.065, 27 ms p50 and 60 ms p95 per decision, about 590 input tokens per decision.
Where it lags: One-pass only, no reasoning: multi-step arithmetic, date differences and documents beyond 8,192 tokens are weak, and calibration is fitted on Kev's suite so other distributions can be over- or under-confident.
What K2-Type-0.9B Is
Questions share the state but cannot see each other (block-causal attention mask), so adding a question never changes another's answer. A pointer head scores each option's </opt> hidden state against the question's <decide> hidden state, and a softmax at temperature 1.478 produces probabilities. Training was a full fine-tune on about 354K decision records (public classification, NLI and QA sets, Kev's decision data, game positions and synthetic items) with soft labels from a 7B teacher decision model, then 50 iterations of PPO on Snake and a temperature refit. The same weights play Snake from a text board (66 food per game on average on a 12x12 board).
Training code and data are not released. Weights and the minimal serving code (jev/) are public. It builds on K2-Horizon-0.9B.
Specifications
| Field | Value |
|---|---|
| Organization | IFM |
| Hugging Face | IFM/K2-Type-0.9B |
| Base model | IFM/K2-Horizon-0.9B |
| Parameters | 0.9B (1.08B stored, including the unused LM head) |
| Output | Typed probabilities (no text generation) |
| License | Apache 2.0 |
| Release date | 2026-09-26 |
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| JevBench (public set, accuracy) | 76.2 | Model card | 2026-09 |
Self-reported by IFM on the 231-item JevBench public set (standard 66/72, easy 47/48, hard 63/111). The official JevBench score also uses sealed items, on which every listed system scores well below its public accuracy.