TR Quantisation - a Ternii Collection

Ternii 's Collections

TR Quantisation

TR Quantisation

updated 2 days ago

Qwen models packed into trits (2.25-3 trits per weight, no retraining) for vLLM and llama.cpp. Runtime: github.com/ternii-ai/tritq-vllm

Ternii/Qwen3.6-35B-A3B-TR2.25-G

Text Generation • 10B • Updated 5 days ago • 425

Note vLLM, one DGX Spark: 180.9 tok/s at 1 user, 17.4 GiB served (NVIDIA NVFP4: 128.0, 21.1 GiB).
Ternii/Qwen3.6-35B-A3B-TR2.25-G-GGUF

Text Generation • 35B • Updated 2 days ago • 583

Note llama.cpp on Mac, 16.5 GB file. Ready-built binaries available.
Ternii/Qwen3.6-27B-TR2.5-G

Text Generation • 10B • Updated 4 days ago • 329 • 1

Note vLLM, dense 27B: 40.5 tok/s at 1 user on one DGX Spark.
Ternii/Qwen3.8-Flash-Next-TR2.57-G

Text Generation • 40B • Updated 4 days ago • 39 • 1

Note vLLM, 125B MoE: 63.8 GiB served (NVIDIA NVFP4: 75.6 GiB).
Ternii/Qwen3.8-Flash-Next-TR3-G-GGUF

Text Generation • 177B • Updated 3 days ago • 240

Note llama.cpp on Mac, experimental, 64 GB class.