---
title: "Dendrite Holdings Squeezes a 27B Model Into 6 GB With Ingot-XVII"
subtitle: "A crowd-sourced ternary quantization of Qwen3.8-27B runs 27B parameters at 25 tokens per second on a base M5 MacBook in 13 GB of RAM."
url: "https://alphasignal.ai/news/dendrite-holdings-squeezes-a-27b-model-into-6-gb-with-ingot-xvii"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-10-30T23:43:59.000Z"
updated: "2026-11-04T18:28:07.717Z"
type: "Model"
topics:
  - "Llms"
  - "Gpus"
subtopics:
  - "Small Models"
  - "Quantization"
upvotes: 27
source_url: "https://huggingface.co/dendriteholdings/Ingot-XVII-27B-Q2"
comments: 0
truncated: true
---

# Dendrite Holdings Squeezes a 27B Model Into 6 GB With Ingot-XVII

> A crowd-sourced ternary quantization of Qwen3.8-27B runs 27B parameters at 25 tokens per second on a base M5 MacBook in 13 GB of RAM.

Dendrite Holdings has published [Ingot-XVII](https://huggingface.co/dendriteholdings/Ingot-XVII-27B-Q2), a quantized Qwen derivative with 27 billion parameters and files as small as 5.95 GB. A Bittensor competition produced the weights by optimizing a ternary model against a fixed BF16 teacher. The release ties that teacher on MATH-500 in one reported run, improves two benchmarks over its ternary starting point, and runs on consumer hardware through a custom llama.cpp fork.

## Subnet 97 makes compression competitive

Bittensor divides work among specialized networks called subnets. On [Subnet 97](https://subnetalpha.ai/subnet/ingot/), miners submit compressed copies of Qwen3.8-27B while a validator measures how closely each copy reproduces the reference model. Submissions must preserve the teacher's architecture and remain within the subnet's size limit.

The subnet previously operated as Albedo, where miners distilled a large model into smaller architectures. Renamed Ingot, it now rewards faithful compression of a fixed model. Ingot-XVII is the seventeenth winning submission under that process.

Its starting weights came from Prism ML's [Bonsai 2](https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf). That model represents each quantized weight with one of three values: {-1, 0, +1}. A separate scale for each block converts those compact codes into effective weight magnitudes during inference.

## Frozen codes, retuned scales

Ternary quantization reduces storage and memory bandwidth by replacing conventional floating-point weights with three-value codes. BitNet-style training exposes a model to simulated quantization errors so that it can tolerate this restricted representation during inference.

Ingot-XVII leaves Bonsai 2's ternary codes intact and adjusts the surrounding parameters:

* None of the 26.87 billion ternary codes changed.
* About 99.8% of the per-block FP16 scales changed.
* The median scale adjustment was 9.6%.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [Ingot-XVII-27B-Q2 model card](https://huggingface.co/dendriteholdings/Ingot-XVII-27B-Q2)
- [Ingot - Subnet Alpha](https://subnetalpha.ai/subnet/ingot/)
- [Dendrite llama.cpp fork](https://github.com/DendriteLabs/llama.cpp)
- [Training low-bit ternary models with Axolotl](https://huggingface.co/blog/axolotl-ai-co/finetuning-ternary-llms-tii-axolotl)
- [Ternary Bonsai 2 27B](https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf)
