Research

Small models.
Real adaptation.

4 Research programs
157M – 8.8B Parameters across the models
6 Public research repositories

How they fit together.

The same question — how much capability can you get on hardware you own? — approached from four directions.

ProgramApproachStarting pointTrained onRuns on
Project Kestrel Design and train a new architecture: hybrid attention, looped core, sparse memory Nothing — ~157M parameters from scratch One rented RTX 3090 for v0.1, now a local RTX 3060 Nearly any computer
Fable-Coder Replay-anchored LoRA post-training for agentic tool use and security Qwen2.5-Coder-7B-Instruct A100 (V4 in ~32 minutes) An RTX 3060 (4.36 GiB GGUF)
Threnody 3.0 Multi-schema LoRA post-training with anti-loop filtering and refusal orthogonalization Qwen/Qwen3.5-9B (~8.8B parameters) A100-80GB (s0.75 merged weights) An RTX 3060 (5.6 GiB Q4_K_M GGUF)
MalxLabs-V4 Supervised fine-tune for math and code reasoning, with aggressive inference tuning DeepSeek-R1-Distill-Qwen-1.5B A rented RTX 4500 Ada A 2014-era GTX 1080

Every claim on these pages links back to a repository, and results are reported the way they came out — including ties, regressions and failed iterations.

Get involved

Help write the evals.

Two one-week, remote internships are open: curating evaluation prompts for Project Kestrel, or turning the research into something a curious non-expert can follow.