# LLaMa Adds Maple 20B-A1B MoE Architecture with Ternary Quantization

Digest AI · Generative AI & Models · published 2026-09-14T12:11:18Z

Canonical: https://digestai.news/story/llama-adds-maple-20b-a1b-moe-architecture-with-ternary-quantization

## Summary

AlexGabbia, a contributor to the LLaMa project, has added the Maple 20B-A1B ternary Multi-Expert (MoE) architecture to the LLaMa codebase. This update includes support for both q20 and tQ10/TQ20 quantization schemes. The team ran tests on a RTX 5070 Ti GPU, finding that TQ20 performed slightly worse than q20 in perplexity scores across different contexts. They also noted differences between CPU and CUDA backends for the same model file, attributing these to backend-specific optimizations rather than quantization or architecture issues. The team is now working on aligning this release with other updates and preparing documentation for the new Maple 20B-A1B MoE architecture.

## Key points

- LLaMa adds support for Maple 20B-A1B ternary MoE architecture
- TQ10/TQ20 quantization scheme included in the update
- CPU and CUDA backends show different performance metrics

## Why it matters

This update is significant as it expands LLaMa's capabilities to support a new type of model architecture, which could be useful for improving efficiency or flexibility in AI applications.

## Sources

1. [llama: add Maple 20B-A1B ternary MoE architecture (CPU) by AlexGabbia · Pull Request #27000 · ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp/pull/27000) (github.com, 2026-09-14, primary source)

## Cite

Digest AI, "LLaMa Adds Maple 20B-A1B MoE Architecture with Ternary Quantization", 14 September 2026, https://digestai.news/story/llama-adds-maple-20b-a1b-moe-architecture-with-ternary-quantization

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/llama-adds-maple-20b-a1b-moe-architecture-with-ternary-quantization.json