K2-Type-0.9B
A 0.9B decision model from the Institute of Foundation Models, K2-Horizon-0.9B fully fine-tuned with a pointer head. A state and any number of noul, choice or score questions go in; one forward pass returns a probability per option, with no generated text.
Choice takes 1 to 255 options and score 2 to 255 levels. A block-causal mask keeps questions independent, and a softmax at a fitted temperature (1.478) gives the probabilities. The repo includes a small server for TypeSafe's /v1/systemone format. 1.08B parameters are stored, including the base model's unused LM head. The maker reports these numbers from its own run of the third-party JevBench public set (231 items): 176 correct (standard 66/72, easy 47/48, hard 63/111), Brier 0.328, ECE 0.065; JevBench's maintainers have queued it but not measured it. Limits the maker states: training code and data are not released; calibration was fitted on Kev's suite, so it may be off on other data; weak at multi-step arithmetic, dates and documents over 8,192 tokens; English only; needs one CUDA GPU and transformers 5.17+ with remote code. First published on 26 Sep as K2-Type-1B and moved into the IFM organisation on 2 Oct.
What it decides
- choice — picks one option from a set
- score — places the input on an ordered scale
- noul — answers a yes/no question with one probability
At a glance
| Parameters | 0.9B |
| Base model | IFM/K2-Horizon-0.9B |
| Maker | Institute of Foundation Models |
| Released | 2026-09-26 |
| License | apache-2.0 |
| Reported accuracy | 76.2% |
| Reported latency | 27 ms p50 / 60 ms p95 per decision on one H200, local, about 590 input tokens per decision |
Get the weights
pip install systemonemodels
systemone pull ifm/k2-type
The files are served from the maker's Hugging Face repository, IFM/K2-Type-0.9B, and verified against the checksums recorded here.