Flagship · Rust · C++ · CUDA
TAK: task-aware model quantization
TAK shrinks large language models to a fixed file size while protecting the skill you care about. Instead of compressing every tensor the same way, it measures how each precision change affects a target task, solves for the best allocation under an exact byte budget, and exports a standard GGUF that runs anywhere llama.cpp does. No fine-tuning, no model merging: only smarter precision.
Qwen3.8-27B · reasoning
82.81 vs 77.34
TAK beat Unsloth's UD-IQ2_S by +5.47 points at the identical 8.37 GB file size, landing within 0.78 points of the 54.7 GB original: 85% smaller for nearly the same reasoning.
Gemma 4 E4B · reasoning
+14.06
points over Unsloth UD-IQ2_M, with a 9% smaller file
Qwen3.5-4B · reasoning
+11.72
points over Unsloth UD-IQ2_M at the same size
- Every score comes from the same frozen 960-item evaluation pack, run to completion with zero failed items. The headline is the run's target category; the other suites are diagnostics.
- Rust application with a C++ worker on a pinned llama.cpp runtime. Runs on Vulkan or CUDA, with resumable, checkpointed jobs and a headless GPU deployment for models too large for a workstation.
- Measures real tensor-format changes against the task's own grader, including executed-test grading for coding, and records signed effects of each change, alone and in pairs.
- Published winners on Hugging Face as the TAK Reasoning Champions collection.
