TernaryQuench: open-source ternary quantization trainer for Qwen3 with MLX export
casper_hansen_ · x · 2026-10-05
Developer penk released TernaryQuench, an open-source project giving the complete recipe for building ternary language models: start from an upstream Qwen3 checkpoint, generate agentic calibration traces, train with CAT-Q-style reconstruction, and export for local inference.
- Supports Qwen3 and the text decoders of Qwen3.5/Qwen3.8; weights learn three values per group (−scale, 0, +scale)
- Exporters let you mix trained ternary layers with higher-precision layers in one model
- Full pipeline open-sourced: calibration, training, checkpoint recovery, evaluation
- Ships a ready TernaryQuench Qwen3.8-27B GGUF runnable via ollama run, with MLX export for Apple Silicon
Related event: TernaryQuench Open-Sources Full Ternary Quantization Recipe for Qwen3(2 posts)→
More from Infra
- New US chip rules reportedly require licenses for non-US fabs and restrict overseas datacenters — JensHonack · 2026-11-21
- Portuguese tutorial walks through running MiniMax H3 locally on PC — Ashamed-Rub5601 · 2026-11-21
- SenseNova Loop 9B on a single RTX 4080: 256k context for 16GB VRAM GPUs — roofkid · 2026-11-21
- Ex-NVIDIA staffer joins Gimlet Labs after NVIDIA's reported $20B bet on Groq's inference tech — yawnxyz · 2026-11-21
- Fully local pipeline: open-weight 27B models ship production code with Playwright + vision-loop tests — awitod · 2026-11-21
- Vespa co-founders launch Hornet: solving retrieval for the agentic era — jobergum · 2026-11-21