typecastlm-qwen3.5-3.8b
Typed decisions from a frozen, cut-down Qwen3.5-4B: the last four blocks are removed and a computed 39-row head reads yes/no, true/false/unsure, one of 2 to 26 options or a level on a scale in one forward pass, with a temperature per mode. Training-free; Apache-2.0.
Per the NOTICE, Qwen3.5-4B blocks 0 to 27 are kept bit for bit and blocks 28 to 31 removed, so it cannot generate text; the head's vectors are computed from the model's own output rows, not trained. Four modes: noul, tfu (adds "unsure"), choice (2 to 26 options, degrading past six) and scale. The answer follows an option's position more than its name, by the maker's own measurement. One temperature per mode, fitted on unnamed broad sets; the maker advises refitting on about 200 labelled rows. Runs as a transformers checkpoint or as llama.cpp GGUFs through the maker's reader; the typecastlm server offers a Jev-compatible HTTP API. Not affiliated with TypeSafe AI.
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 | 3.76B |
| Base model | Qwen/Qwen3.5-4B |
| Maker | Mikhail Gribov |
| Released | 2026-09-24 |
| License | apache-2.0 |
| Reported accuracy | 85.6% |
| Reported latency | 48 ms p50 / 75 ms p90 for inputs under 200 tokens on a 16 GB consumer GPU |
Get the weights
pip install systemonemodels
systemone pull mikhail-gribov/typecastlm
The files are served from the maker's Hugging Face repository, mihailgribov/typecastlm-qwen3.5-3.8b, and verified against the checksums recorded here.
Read more
This page was opened by System One for Mikhail Gribov, who can claim the organisation and take it over at any time.