CMSManhattan/JiRackTernaryPrecision-72B-chat

JiRack Ternary Precision 72B (GGUF)

JiRack Ternary Precision 72B is a ternary (1.58-bit) language model by CMSManhattan (Konstantin Grabko) with 72.15B parameters, the largest model of the JiRack Precision line, built for coding, tool calling and agents. Every linear layer of the transformer uses ternary weights {-1, 0, +1} with a per-tensor scale and 8-bit activations (BitNet b1.58 style), trained with quantization-aware training.

This repository has the model as GGUF for llama.cpp: about @SIZE_TQ2@ instead of 144.3 GB in bf16. It runs on a CPU server with enough RAM, or on one GPU.

It is in development yet

Files

File	Size	Notes
JiRackTernaryPrecision-72B-TQ2_0.gguf	@SIZE_TQ2@	recommended, ternary weights stored natively
JiRackTernaryPrecision-72B-Q8_0.gguf	@SIZE_Q8@	for llama.cpp builds without TQ2_0 kernels on your hardware
chat_template.jinja	—	JiRack chat format, including tools

Run with llama.cpp

llama-server -m JiRackTernaryPrecision-72B-TQ2_0.gguf -c 16384 --jinja --port 8080

The server is OpenAI-compatible (http://localhost:8080/v1/chat/completions). On a GPU add -ngl 999 to put all layers on it; the TQ2_0 file needs about @SIZE_TQ2@ of memory plus the context cache.

Tool calling

Pass tools=[...] in OpenAI format. The model calls tools with one or more blocks:

<|tool_call_start|>
{"name": "get_weather", "arguments": {"city": "Paris"}}
<|tool_call_end|>

Several blocks in one answer are parallel calls. Tool results go back as role: "tool" messages. When no tool fits, the model answers in plain text; when a required argument is missing, it asks for it.

Model details

Parameters	72.15B
Layers	132
Hidden size / FFN	5,120 / 30,720
Attention heads	40 query, 8 key-value (GQA), head dim 128
Vocabulary	152,064
Context	up to 16k tokens
Weights	ternary {-1, 0, +1}, per-tensor scale
Activations	8-bit, per token

How the model was made

The 72B grows out of JiRack Ternary Precision 36B. The base CMSManhattan/JiRackUltra_32b got a wider FFN (27,648 → 30,720, the 36B), then more depth: 132 layers, the 64 layers of the 36B plus 68 inserted layers. The inserted layers start as exact identity blocks, so the 72B first answers exactly like the 36B; training then puts the new capacity to work.

Intended use

Coding assistants, tool calling and agent pipelines where quality matters more than speed, run on a single server instead of a GPU cluster.

Limitations

The model can make mistakes, invent facts and produce wrong tool arguments. Check its output before acting on it, especially in automated pipelines.

License

Proprietary, © 2025–2026 Konstantin Vladimirovich Grabko (CMSManhattan).

Contact

CMSManhattan — Konstantin Grabko, Hugging Face: kgrabko