Run Kate 27B (abliterated) locally on a 12 to 32 GB GPU
Kate is Katy (Swift-1.5-Qwen3.8-27B) with the refusal direction removed from the weights, as IQ4_XS in bartowski's tensor layout with our own importance matrix, plus the public DSpine b16 draft head. Requests that only sound unsafe are no longer refused (OR-Bench hard: 88 of 100 for Katy, 1 for Kate); clearly harmful ones still are, 70 of 100.
flelink run kate-27b-dspineWhat you need
One NVIDIA card from 12 GB and 24 GB of RAM, and FLE v1.3.0 (the dense engine line), which FLE Link installs for this recipe.
Measured speed
| Recipe | Decode | Measured on |
|---|---|---|
| Kate 27B + DSpineKaty without the refusal direction | 142.7 tok/s | from 20 GB up all 66 layers are on the card; memory limit simulated on a 32 GB card, FLE v1.3.0 source, Nov 2026 |
| Kate 27B + DSpineKaty without the refusal direction | 37.3 tok/s | 16 GB, 55 of 66 layers on the card; simulated on a 32 GB card in a Ryzen 9 9950X3D with 128 GB RAM; a slower CPU will be slower |
| Kate 27B + DSpineKaty without the refusal direction | 14.1 tok/s | 12 GB, 37 of 66 layers; memory limited on a larger card to simulate the size |
Does Kate 27B + DSpine fit my card?
| Graphics | 32 GB RAM | 64 GB RAM | 128 GB RAM |
|---|---|---|---|
| 12 GB NVIDIA card | fits | fits | fits |
| 16 GB NVIDIA card | fits | fits | fits |
| 24 GB NVIDIA card | fits | fits | fits |
| 32 GB NVIDIA card | fits | fits | fits |
| Apple Silicon Mac | no engine | no engine | no engine |
| No GPU | no engine | no engine | no engine |
"no engine": this recipe needs the FLE engine, which has an NVIDIA CUDA build only. File size 16 GiB including the draft model.
The verdicts are the rule FLE Link applies before it downloads anything (the rules in full). A speed is shown only for the hardware it was measured on. A size nobody measured yet? flelink bench --task pagoda --share posts yours to the benchmarks board.
Good to know
- Kate no longer refuses requests the original model refused for safety reasons. You are responsible for what you do with it.
- Quality against Katy on the same build: ARC-Challenge 94.2 % (Katy 94.3), GSM8K 97.6 % (97.2), LiveCodeBench 44 (42); no difference is significant.
- Below 20 GB the speed depends on the CPU and the RAM: measured on a Ryzen 9 9950X3D with 128 GB RAM, with the memory limit simulated on a 32 GB card.
- Licence of the model: Swift Open License v1.0, the DSpine head Apache-2.0, see the model page.
Sources
Install FLE Link
curl -fsSL https://flelink.run/install.sh | shWindows: irm https://flelink.run/install.ps1 | iex. Open source, MIT. No account, no telemetry. What FLE Link does →