kestrel/kestrel-0.5b-draft
Text generation500M paramssafetensorsLicense: apache-2.0en#speculative-decoding#draft
Updated Aug 26, 2026
Part of the Open Pace seed catalog. Weights are listed for the preview and become downloadable at launch.
Overview
A draft model for speculative decoding with the Kestrel 7B line.
kestrel/kestrel-0.5b-draft is a 500M-parameter text generation model, published in safetensors under the apache-2.0 license.
Intended use
- Text generation in en.
- Research, prototypes and products that keep a person in the loop.
- Fine-tuning as a starting point for a narrower task.
How to use
# Command-line client (planned; the shape may change)
pace pull kestrel/kestrel-0.5b-draft
# Python (planned)
from openpace import load
model = load("kestrel/kestrel-0.5b-draft")Limitations
- It can state wrong things fluently. Check facts that matter.
- Quality drops on languages and domains that were thin in its training data.
- It reflects biases present in public text.
License
Released under apache-2.0. Read the license file before using the weights commercially.
More text generation
kestrel/kestrel-1.5b-base
The pretrained base behind the Kestrel instruct line, released for anyone who wants to tune their own.
Text generation1.5BUpdated Sep 20, 2026
byte-gardener/tinystories-26m
A 26-million-parameter model trained only on short children's stories. Useful for teaching.
Text generation26MUpdated Sep 12, 2026
byte-gardener/tinystories-110m
A larger run on the same story corpus, to compare what scale adds.
Text generation110MUpdated Aug 20, 2026