I'm Kyle Hessling. I build, merge, and evaluate open language models from a two-GPU lab in Wyoming, and publish everything. My releases have been downloaded more than a million times. Digital Outlook is where I take that work to businesses.
Illustration: a wide mountain range drawn as contour lines, representing a large teacher model. As you scroll, it erodes into one sharp peak, a small specialist model that keeps the teacher's ridge shape. The summit contour turns amber.
Most of a giant model sits idle on any single project. I think the next gains for people running models at home come from specialists: smaller models trained on how much larger models behave in one domain, sized to fit on one GPU. Load the specialist you need, when you need it.
I call this the Single-GPU Frontier, and it's what I work on. Two workstreams: distilling domain experts from much larger open teachers, and teaching models to reason without wandering. Both run on two devices I own.
Every row says what I did on it. Two are mine; the rest are collaborations, credited to the people who led them.
27B dense502,598 all-time downloads
A code-heavy finetune had broken the model’s chain of thought. Fusion takes the difference between the coder and its stable parent as a task vector and merges it back in, a little at the bottom of the network and more toward the top. Reasoning stays intact and the coding gains carry over.
18B743,796 all-time downloads
A frankenmerge: transformer layers stacked into a new 18B model, then healed with QLoRA so the seams stop showing. The frontier models I asked advised against it. It became one of my most downloaded releases, and the full recipe is published step by step.
27B dense282,043 all-time downloads
Qwen 3.8 27B tuned to think less wastefully in agent loops. SFT on the best 10% of about 1.5M teacher examples, then NeMo-RL with GSPO. We cut runaway reasoning and sped up decoding, and paid a small price on MMLU-Pro that the card publishes.
Tradeoff, published on the card: MMLU-Pro mixed set: 91.28% vs 92.73% for the base (−1.45 points).
27B dense and 35B-A3B MoE
The Qwopus 3.6 line, led by Jackrong. I contributed training compute, data curation, a tool-calling chat template that works across runtimes, agentic demo Spaces, and independent evaluations, and I’m credited on the cards for hardware infrastructure, training support, and live agent experiments.
Private, specialized models that run on hardware you own, and proof they work before anyone commits.
Apps, ecommerce, and the integrations that connect old systems to new ones.
